Cherry-Pick from 16662 : Anakin subgraph cpu support

cmake/anakin_subgraph.cmake

@@ -25,8 +25,9 @@ endif()
 
 if(ANAKIN_FOUND)
     message(STATUS "Current ANAKIN header is ${ANAKIN_INCLUDE_DIR}/anakin_config.h. ")
+    include_directories(${ANAKIN_ROOT})
     include_directories(${ANAKIN_ROOT}/include)
-    include_directories(${ANAKIN_ROOT}/include/saber)
+    include_directories(${ANAKIN_ROOT}/saber)
     link_directories(${ANAKIN_ROOT})
     add_definitions(-DPADDLE_WITH_ANAKIN)
 endif()

paddle/fluid/inference/anakin/convert/activation.cc

@@ -16,16 +16,13 @@
 #include <algorithm>
 #include <map>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::saber::NV;
-using anakin::saber::Shape;
-
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-ActivationOpConverter::ActivationOpConverter(const std::string &op_type)
+template <typename TargetT>
+ActivationOpConverter<TargetT>::ActivationOpConverter(
+    const std::string &op_type)
     : op_type_(op_type) {
   auto it = anakin_op_types_.find(op_type_);
   PADDLE_ENFORCE(it != anakin_op_types_.end(),
@@ -33,10 +30,10 @@ ActivationOpConverter::ActivationOpConverter(const std::string &op_type)
   anakin_op_type_ = it->second;
 }
 
-void ActivationOpConverter::operator()(const framework::proto::OpDesc &op,
-                                       const framework::BlockDesc &block_desc,
-                                       const framework::Scope &scope,
-                                       bool test_mode) {
+template <typename TargetT>
+void ActivationOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
   PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
@@ -44,13 +41,20 @@ void ActivationOpConverter::operator()(const framework::proto::OpDesc &op,
   auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
   auto input_name = op_desc.Input("X").front();
   auto output_name = op_desc.Output("Out").front();
-  engine_->AddOp(op_name, "Activation", {input_name}, {output_name});
-  engine_->AddOpAttr(op_name, "type", anakin_op_type_);
+  this->engine_->AddOp(op_name, "Activation", {input_name}, {output_name});
+  this->engine_->AddOpAttr(op_name, "type", anakin_op_type_);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(sigmoid, SigmoidOpConverter);
-REGISTER_ANAKIN_OP_CONVERTER(tanh, TanhOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(sigmoid,
+                                  SigmoidOpConverter<::anakin::saber::NV>);
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(tanh, TanhOpConverter<::anakin::saber::NV>);
+#endif
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(sigmoid,
+                                 SigmoidOpConverter<::anakin::saber::X86>);
+REGISTER_CPU_ANAKIN_OP_CONVERTER(tanh, TanhOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/activation.h

@@ -22,7 +22,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class ActivationOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class ActivationOpConverter : public AnakinOpConverter<TargetT> {
  public:
   explicit ActivationOpConverter(const std::string &op_type);
 
@@ -39,14 +40,16 @@ class ActivationOpConverter : public AnakinOpConverter {
                                                       {"sigmoid", "Sigmoid"}};
 };
 
-class TanhOpConverter : public ActivationOpConverter {
+template <typename TargetT>
+class TanhOpConverter : public ActivationOpConverter<TargetT> {
  public:
-  TanhOpConverter() : ActivationOpConverter("tanh") {}
+  TanhOpConverter() : ActivationOpConverter<TargetT>("tanh") {}
 };
 
-class SigmoidOpConverter : public ActivationOpConverter {
+template <typename TargetT>
+class SigmoidOpConverter : public ActivationOpConverter<TargetT> {
  public:
-  SigmoidOpConverter() : ActivationOpConverter("sigmoid") {}
+  SigmoidOpConverter() : ActivationOpConverter<TargetT>("sigmoid") {}
 };
 }  // namespace anakin
 }  // namespace inference

paddle/fluid/inference/anakin/convert/affine_channel.cc

@@ -18,19 +18,16 @@
 #include <vector>
 
 using anakin::graph::GraphGlobalMem;
+using anakin::PTuple;
 using anakin::AK_FLOAT;
-using anakin::Precision;
-using anakin::saber::NV;
-using anakin::saber::X86;
 using anakin::saber::Shape;
-using anakin::PBlock;
-using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void AffineChannelOpConverter::operator()(
+template <typename TargetT>
+void AffineChannelOpConverter<TargetT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -59,7 +56,7 @@ void AffineChannelOpConverter::operator()(
   bias_tensor->Resize(bias_t->dims());
   TensorCopySync((*bias_t), platform::CPUPlace(), bias_tensor.get());
 
-  engine_->AddOp(op_name, "AffineChannel", {input_name}, {output_name});
+  this->engine_->AddOp(op_name, "AffineChannel", {input_name}, {output_name});
 
   // Generate the Scale parameter of Anakin.
   auto scale_shape = framework::vectorize2int(scale_t->dims());
@@ -67,15 +64,16 @@ void AffineChannelOpConverter::operator()(
     scale_shape.insert(scale_shape.begin(), 1);
   }
   Shape anakin_scale_shape(scale_shape);
-  auto *weight1 = GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(
-      anakin_scale_shape);
+  auto *weight1 =
+      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
+          anakin_scale_shape);
   float *scale_cpu_data =
       static_cast<float *>(weight1->h_tensor().mutable_data());
   std::copy_n(scale_tensor->data<float>(), scale_tensor->numel(),
               scale_cpu_data);
   weight1->d_tensor().set_shape(anakin_scale_shape);
   weight1->d_tensor().copy_from(weight1->h_tensor());
-  engine_->AddOpAttr(op_name, "weight_1", *weight1);
+  this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
 
   // Generate the Bias parameter of Anakin.
   auto bias_shape = framework::vectorize2int(bias_t->dims());
@@ -83,18 +81,24 @@ void AffineChannelOpConverter::operator()(
     bias_shape.insert(bias_shape.begin(), 1);
   }
   Shape anakin_bias_shape(bias_shape);
-  auto *weight2 = GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(
-      anakin_bias_shape);
+  auto *weight2 =
+      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
+          anakin_bias_shape);
   float *bias_cpu_data =
       static_cast<float *>(weight2->h_tensor().mutable_data());
   std::copy_n(bias_tensor->data<float>(), bias_tensor->numel(), bias_cpu_data);
   weight2->d_tensor().set_shape(anakin_bias_shape);
   weight2->d_tensor().copy_from(weight2->h_tensor());
-  engine_->AddOpAttr(op_name, "weight_2", *weight2);
+  this->engine_->AddOpAttr(op_name, "weight_2", *weight2);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(affine_channel, AffineChannelOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(
+    affine_channel, AffineChannelOpConverter<::anakin::saber::NV>);
+#endif
+REGISTER_CPU_ANAKIN_OP_CONVERTER(
+    affine_channel, AffineChannelOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/affine_channel.h

@@ -21,7 +21,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class AffineChannelOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class AffineChannelOpConverter : public AnakinOpConverter<TargetT> {
  public:
   AffineChannelOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/batch_norm.cc

@@ -21,17 +21,16 @@
 
 using anakin::graph::GraphGlobalMem;
 using anakin::AK_FLOAT;
-using anakin::saber::NV;
 using anakin::saber::Shape;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void BatchNormOpConverter::operator()(const framework::proto::OpDesc &op,
-                                      const framework::BlockDesc &block_desc,
-                                      const framework::Scope &scope,
-                                      bool test_mode) {
+template <typename TargetT>
+void BatchNormOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Output("Y").size(), 1);
   std::map<std::string, std::string> inputs;
@@ -48,9 +47,9 @@ void BatchNormOpConverter::operator()(const framework::proto::OpDesc &op,
 
   auto bn_op_name = op_name + ":bn";
   auto bn_output = bn_op_name + "_output";
-  engine_->AddOp(bn_op_name, "BatchNorm", {inputs["X"]}, {bn_output});
-  engine_->AddOpAttr(bn_op_name, "epsilon", epsilon);
-  engine_->AddOpAttr(bn_op_name, "momentum", static_cast<float>(1.0));
+  this->engine_->AddOp(bn_op_name, "BatchNorm", {inputs["X"]}, {bn_output});
+  this->engine_->AddOpAttr(bn_op_name, "epsilon", epsilon);
+  this->engine_->AddOpAttr(bn_op_name, "momentum", static_cast<float>(1.0));
 
   auto scale_op_name = op_name + ":scale";
   auto get_lod_tensor = [this, &scope, &op_name](const std::string &var_name,
@@ -81,48 +80,54 @@ void BatchNormOpConverter::operator()(const framework::proto::OpDesc &op,
   Shape shape1(fill_shape(4, framework::vectorize2int(mean_t.dims())));
   Shape shape2(fill_shape(4, framework::vectorize2int(variance_t.dims())));
   auto *weight1 =
-      GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(shape1);
+      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(shape1);
   auto *mean_data = static_cast<float *>(weight1->h_tensor().mutable_data());
   std::copy_n(mean_t.data<float>(), mean_t.numel(), mean_data);
-  engine_->AddOpAttr(bn_op_name, "weight_1", *weight1);
+  this->engine_->AddOpAttr(bn_op_name, "weight_1", *weight1);
 
   auto *weight2 =
-      GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(shape2);
+      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(shape2);
   auto *variance_data =
       static_cast<float *>(weight2->h_tensor().mutable_data());
   std::copy_n(variance_t.data<float>(), variance_t.numel(), variance_data);
-  engine_->AddOpAttr(bn_op_name, "weight_2", *weight2);
+  this->engine_->AddOpAttr(bn_op_name, "weight_2", *weight2);
 
   Shape shape3(std::vector<int>({1, 1, 1, 1}));
   auto *weight3 =
-      GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(shape3);
+      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(shape3);
   auto *alpha_data = static_cast<float *>(weight3->h_tensor().mutable_data());
   float weight3_data[] = {1};
   std::copy(std::begin(weight3_data), std::end(weight3_data), alpha_data);
-  engine_->AddOpAttr(bn_op_name, "weight_3", *weight3);
+  this->engine_->AddOpAttr(bn_op_name, "weight_3", *weight3);
 
   Shape scale_shape(fill_shape(4, framework::vectorize2int(scale_t.dims())));
-  auto *scale =
-      GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(scale_shape);
+  auto *scale = GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
+      scale_shape);
   auto *scale_data = static_cast<float *>(scale->h_tensor().mutable_data());
   std::copy_n(scale_t.data<float>(), scale_t.numel(), scale_data);
 
   Shape bias_shape(fill_shape(4, framework::vectorize2int(bias_t.dims())));
-  auto *bias =
-      GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(bias_shape);
+  auto *bias = GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
+      bias_shape);
   auto *bias_data = static_cast<float *>(bias->h_tensor().mutable_data());
   std::copy_n(bias_t.data<float>(), bias_t.numel(), bias_data);
 
-  engine_->AddOp(scale_op_name, "Scale", {bn_output}, {output});
-  engine_->AddOpAttr(scale_op_name, "axis", 1);
-  engine_->AddOpAttr(scale_op_name, "num_axes", 1);
-  engine_->AddOpAttr(scale_op_name, "bias_term", true);
-  engine_->AddOpAttr(scale_op_name, "weight_1", *scale);
-  engine_->AddOpAttr(scale_op_name, "weight_2", *bias);
+  this->engine_->AddOp(scale_op_name, "Scale", {bn_output}, {output});
+  this->engine_->AddOpAttr(scale_op_name, "axis", 1);
+  this->engine_->AddOpAttr(scale_op_name, "num_axes", 1);
+  this->engine_->AddOpAttr(scale_op_name, "bias_term", true);
+  this->engine_->AddOpAttr(scale_op_name, "weight_1", *scale);
+  this->engine_->AddOpAttr(scale_op_name, "weight_2", *bias);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(batch_norm, BatchNormOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(batch_norm,
+                                  BatchNormOpConverter<::anakin::saber::NV>);
+#endif
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(batch_norm,
+                                 BatchNormOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/batch_norm.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class BatchNormOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class BatchNormOpConverter : public AnakinOpConverter<TargetT> {
  public:
   BatchNormOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/concat.cc

@@ -15,38 +15,32 @@
 #include "paddle/fluid/inference/anakin/convert/concat.h"
 #include <algorithm>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::Precision;
-using anakin::saber::NV;
-using anakin::saber::X86;
-using anakin::saber::Shape;
-using anakin::PBlock;
-using anakin::PTuple;
-
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void ConcatOpConverter::operator()(const framework::proto::OpDesc &op,
-                                   const framework::BlockDesc &block_desc,
-                                   const framework::Scope &scope,
-                                   bool test_mode) {
+template <typename TargetT>
+void ConcatOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   int axis = boost::get<int>(op_desc.GetAttr("axis"));
   auto input_names = op_desc.Input("X");
-  // PADDLE_ENFORCE(axis > 0,
-  //               "The axis attr of Concat op should be large than 0 for trt");
 
   auto y_name = op_desc.Output("Out").front();
   auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
 
-  engine_->AddOp(op_name, "Concat", input_names, {y_name});
-  engine_->AddOpAttr(op_name, "axis", axis);
+  this->engine_->AddOp(op_name, "Concat", input_names, {y_name});
+  this->engine_->AddOpAttr(op_name, "axis", axis);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(concat, ConcatOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(concat,
+                                  ConcatOpConverter<::anakin::saber::NV>);
+#endif
+REGISTER_CPU_ANAKIN_OP_CONVERTER(concat,
+                                 ConcatOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/concat.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class ConcatOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class ConcatOpConverter : public AnakinOpConverter<TargetT> {
  public:
   ConcatOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/conv2d.cc

@@ -18,19 +18,18 @@
 #include <vector>
 
 using anakin::graph::GraphGlobalMem;
+using anakin::PTuple;
 using anakin::AK_FLOAT;
-using anakin::saber::NV;
 using anakin::saber::Shape;
-using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void Conv2dOpConverter::operator()(const framework::proto::OpDesc &op,
-                                   const framework::BlockDesc &block_desc,
-                                   const framework::Scope &scope,
-                                   bool test_mode) {
+template <typename TargetT>
+void Conv2dOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("Input").size(), 1UL);
   PADDLE_ENFORCE_EQ(op_desc.Input("Filter").size(), 1UL);
@@ -39,7 +38,7 @@ void Conv2dOpConverter::operator()(const framework::proto::OpDesc &op,
   auto input_name = op_desc.Input("Input").front();
   auto output_name = op_desc.Output("Output").front();
   auto op_name = op_desc.Type() + ":" + op_desc.Output("Output").front();
-  engine_->AddOp(op_name, "Convolution", {input_name}, {output_name});
+  this->engine_->AddOp(op_name, "Convolution", {input_name}, {output_name});
 
   auto *filter_v = scope.FindVar(op_desc.Input("Filter").front());
   PADDLE_ENFORCE_NOT_NULL(filter_v);
@@ -51,38 +50,44 @@ void Conv2dOpConverter::operator()(const framework::proto::OpDesc &op,
 
   PADDLE_ENFORCE_EQ(weight_tensor->dims().size(), 4UL);
 
-  // const int n_output = weight_tensor->dims()[0];
-  // const int n_input = weight_tensor->dims()[1];
   const int filter_h = weight_tensor->dims()[2];
   const int filter_w = weight_tensor->dims()[3];
-  // auto filter_num = n_input * filter_h * filter_w ;
+
   auto filter_num = weight_tensor->dims()[0];
-  engine_->AddOpAttr<int>(op_name, "filter_num", filter_num);
-  engine_->AddOpAttr<PTuple<int>>(op_name, "kernel_size", {filter_h, filter_w});
+  this->engine_->template AddOpAttr<int>(op_name, "filter_num", filter_num);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "kernel_size",
+                                                 {filter_h, filter_w});
   auto strides = boost::get<std::vector<int>>(op_desc.GetAttr("strides"));
-  engine_->AddOpAttr<PTuple<int>>(op_name, "strides", strides);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "strides", strides);
   auto paddings = boost::get<std::vector<int>>(op_desc.GetAttr("paddings"));
-  engine_->AddOpAttr<PTuple<int>>(op_name, "padding", paddings);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "padding", paddings);
   auto dilations = boost::get<std::vector<int>>(op_desc.GetAttr("dilations"));
-  engine_->AddOpAttr<PTuple<int>>(op_name, "dilation_rate", dilations);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "dilation_rate",
+                                                 dilations);
   const int groups = boost::get<int>(op_desc.GetAttr("groups"));
-  engine_->AddOpAttr(op_name, "group", groups);
-  engine_->AddOpAttr(op_name, "axis", 1);
-  engine_->AddOpAttr(op_name, "bias_term", false);
+  this->engine_->AddOpAttr(op_name, "group", groups);
+  this->engine_->AddOpAttr(op_name, "axis", 1);
+  this->engine_->AddOpAttr(op_name, "bias_term", false);
 
   auto weight_shape = framework::vectorize2int(filter_t->dims());
   Shape anakin_shape(weight_shape);
   auto *weight1 =
-      GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(anakin_shape);
+      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
+          anakin_shape);
   float *cpu_data = static_cast<float *>(weight1->h_tensor().mutable_data());
   std::copy_n(weight_tensor->data<float>(), weight_tensor->numel(), cpu_data);
   weight1->d_tensor().set_shape(anakin_shape);
   weight1->d_tensor().copy_from(weight1->h_tensor());
-  engine_->AddOpAttr(op_name, "weight_1", *weight1);
+  this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(conv2d, Conv2dOpConverter);
+REGISTER_CPU_ANAKIN_OP_CONVERTER(conv2d,
+                                 Conv2dOpConverter<::anakin::saber::X86>);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(conv2d,
+                                  Conv2dOpConverter<::anakin::saber::NV>);
+#endif

paddle/fluid/inference/anakin/convert/conv2d.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class Conv2dOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class Conv2dOpConverter : public AnakinOpConverter<TargetT> {
  public:
   Conv2dOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/conv2d_fusion.cc

@@ -18,19 +18,18 @@
 #include <vector>
 
 using anakin::graph::GraphGlobalMem;
+using anakin::PTuple;
 using anakin::AK_FLOAT;
-using anakin::saber::NV;
 using anakin::saber::Shape;
-using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void Conv2dFusionOpConverter::operator()(const framework::proto::OpDesc &op,
-                                         const framework::BlockDesc &block_desc,
-                                         const framework::Scope &scope,
-                                         bool test_mode) {
+template <typename TargetT>
+void Conv2dFusionOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("Input").size(), 1UL);
   PADDLE_ENFORCE_EQ(op_desc.Input("Filter").size(), 1UL);
@@ -40,7 +39,7 @@ void Conv2dFusionOpConverter::operator()(const framework::proto::OpDesc &op,
   auto input_name = op_desc.Input("Input").front();
   auto output_name = op_desc.Output("Output").front();
   auto op_name = op_desc.Type() + ":" + op_desc.Output("Output").front();
-  engine_->AddOp(op_name, "Convolution", {input_name}, {output_name});
+  this->engine_->AddOp(op_name, "Convolution", {input_name}, {output_name});
 
   auto *filter_v = scope.FindVar(op_desc.Input("Filter").front());
   PADDLE_ENFORCE_NOT_NULL(filter_v);
@@ -63,28 +62,31 @@ void Conv2dFusionOpConverter::operator()(const framework::proto::OpDesc &op,
   const int filter_w = weight_tensor->dims()[3];
   // auto filter_num = n_input * filter_h * filter_w ;
   auto filter_num = weight_tensor->dims()[0];
-  engine_->AddOpAttr<int>(op_name, "filter_num", filter_num);
-  engine_->AddOpAttr<PTuple<int>>(op_name, "kernel_size", {filter_h, filter_w});
+  this->engine_->template AddOpAttr<int>(op_name, "filter_num", filter_num);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "kernel_size",
+                                                 {filter_h, filter_w});
   auto strides = boost::get<std::vector<int>>(op_desc.GetAttr("strides"));
-  engine_->AddOpAttr<PTuple<int>>(op_name, "strides", strides);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "strides", strides);
   auto paddings = boost::get<std::vector<int>>(op_desc.GetAttr("paddings"));
-  engine_->AddOpAttr<PTuple<int>>(op_name, "padding", paddings);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "padding", paddings);
   auto dilations = boost::get<std::vector<int>>(op_desc.GetAttr("dilations"));
-  engine_->AddOpAttr<PTuple<int>>(op_name, "dilation_rate", dilations);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "dilation_rate",
+                                                 dilations);
   const int groups = boost::get<int>(op_desc.GetAttr("groups"));
-  engine_->AddOpAttr(op_name, "group", groups);
-  engine_->AddOpAttr(op_name, "axis", 1);
-  engine_->AddOpAttr(op_name, "bias_term", true);
+  this->engine_->AddOpAttr(op_name, "group", groups);
+  this->engine_->AddOpAttr(op_name, "axis", 1);
+  this->engine_->AddOpAttr(op_name, "bias_term", true);
 
   auto weight_shape = framework::vectorize2int(filter_t->dims());
   Shape anakin_shape(weight_shape);
   auto *weight1 =
-      GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(anakin_shape);
+      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
+          anakin_shape);
   float *cpu_data = static_cast<float *>(weight1->h_tensor().mutable_data());
   std::copy_n(weight_tensor->data<float>(), weight_tensor->numel(), cpu_data);
   weight1->d_tensor().set_shape(anakin_shape);
   weight1->d_tensor().copy_from(weight1->h_tensor());
-  engine_->AddOpAttr(op_name, "weight_1", *weight1);
+  this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
 
   auto bias_shape = framework::vectorize2int(b_t->dims());
   framework::LoDTensor bias_tensor;
@@ -98,17 +100,24 @@ void Conv2dFusionOpConverter::operator()(const framework::proto::OpDesc &op,
   // bias_shape.push_back(1);
   Shape anakin_bias_shape(bias_shape);
 
-  auto *weight2 = GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(
-      anakin_bias_shape);
+  auto *weight2 =
+      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
+          anakin_bias_shape);
   float *cpu_data2 = static_cast<float *>(weight2->h_tensor().mutable_data());
   std::copy_n(bias_data, bias_tensor.numel(), cpu_data2);
   weight2->d_tensor().set_shape(anakin_bias_shape);
   weight2->d_tensor().copy_from(weight2->h_tensor());
-  engine_->AddOpAttr(op_name, "weight_2", *weight2);
+  this->engine_->AddOpAttr(op_name, "weight_2", *weight2);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(conv2d_fusion, Conv2dFusionOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(conv2d_fusion,
+                                  Conv2dFusionOpConverter<::anakin::saber::NV>);
+#endif
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(conv2d_fusion,
+                                 Conv2dFusionOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/conv2d_fusion.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class Conv2dFusionOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class Conv2dFusionOpConverter : public AnakinOpConverter<TargetT> {
  public:
   Conv2dFusionOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/density_prior_box.cc

@@ -17,17 +17,14 @@
 #include <map>
 #include <vector>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::saber::NV;
-using anakin::saber::Shape;
 using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void DensityPriorBoxOpConverter::operator()(
+template <typename TargetT>
+void DensityPriorBoxOpConverter<TargetT>::operator()(
     const framework::proto::OpDesc& op, const framework::BlockDesc& block_desc,
     const framework::Scope& scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -81,27 +78,44 @@ void DensityPriorBoxOpConverter::operator()(
 
   std::vector<float> temp_v = {};
 
-  engine_->AddOp(op_name, "PriorBox", {input_name, image_name}, {output_name});
-  engine_->AddOpAttr<PTuple<float>>(op_name, "min_size", min_sizes);
-  engine_->AddOpAttr<PTuple<float>>(op_name, "max_size", max_sizes);
-  engine_->AddOpAttr<PTuple<float>>(op_name, "aspect_ratio", aspect_ratios);
-  engine_->AddOpAttr<PTuple<float>>(op_name, "fixed_size", fixed_sizes);
-  engine_->AddOpAttr<PTuple<float>>(op_name, "fixed_ratio", fixed_ratios);
-  engine_->AddOpAttr<PTuple<float>>(op_name, "density", dens);
-  engine_->AddOpAttr(op_name, "is_flip", is_flip);
-  engine_->AddOpAttr(op_name, "is_clip", is_clip);
-  engine_->AddOpAttr<PTuple<float>>(op_name, "variance", variances);
-  engine_->AddOpAttr(op_name, "img_h", static_cast<int>(0));
-  engine_->AddOpAttr(op_name, "img_w", static_cast<int>(0));
-  engine_->AddOpAttr(op_name, "step_h", step_h);
-  engine_->AddOpAttr(op_name, "step_w", step_w);
-  engine_->AddOpAttr(op_name, "offset", offset);
-  engine_->AddOpAttr<PTuple<std::string>>(op_name, "order", t_order);
+  this->engine_->AddOp(op_name, "PriorBox", {input_name, image_name},
+                       {output_name});
+  this->engine_->template AddOpAttr<PTuple<float>>(op_name, "min_size",
+                                                   min_sizes);
+  this->engine_->template AddOpAttr<PTuple<float>>(op_name, "max_size",
+                                                   max_sizes);
+  this->engine_->template AddOpAttr<PTuple<float>>(op_name, "aspect_ratio",
+                                                   aspect_ratios);
+  this->engine_->template AddOpAttr<PTuple<float>>(op_name, "fixed_size",
+                                                   fixed_sizes);
+  this->engine_->template AddOpAttr<PTuple<float>>(op_name, "fixed_ratio",
+                                                   fixed_ratios);
+  this->engine_->template AddOpAttr<PTuple<float>>(op_name, "density", dens);
+  this->engine_->AddOpAttr(op_name, "is_flip", is_flip);
+  this->engine_->AddOpAttr(op_name, "is_clip", is_clip);
+  this->engine_->template AddOpAttr<PTuple<float>>(op_name, "variance",
+                                                   variances);
+  this->engine_->AddOpAttr(op_name, "img_h", static_cast<int>(0));
+  this->engine_->AddOpAttr(op_name, "img_w", static_cast<int>(0));
+  this->engine_->AddOpAttr(op_name, "step_h", step_h);
+  this->engine_->AddOpAttr(op_name, "step_w", step_w);
+  this->engine_->AddOpAttr(op_name, "offset", offset);
+  this->engine_->template AddOpAttr<PTuple<std::string>>(op_name, "order",
+                                                         t_order);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(density_prior_box, DensityPriorBoxOpConverter);
-REGISTER_ANAKIN_OP_CONVERTER(prior_box, DensityPriorBoxOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(
+    density_prior_box, DensityPriorBoxOpConverter<::anakin::saber::NV>);
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(
+    prior_box, DensityPriorBoxOpConverter<::anakin::saber::NV>);
+#endif
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(
+    density_prior_box, DensityPriorBoxOpConverter<::anakin::saber::X86>);
+REGISTER_CPU_ANAKIN_OP_CONVERTER(
+    prior_box, DensityPriorBoxOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/density_prior_box.h

@@ -22,7 +22,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class DensityPriorBoxOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class DensityPriorBoxOpConverter : public AnakinOpConverter<TargetT> {
  public:
   DensityPriorBoxOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/detection_out.cc

@@ -16,19 +16,14 @@
 #include <algorithm>
 #include <map>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::saber::NV;
-using anakin::saber::Shape;
-
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void DetectionOutOpConverter::operator()(const framework::proto::OpDesc &op,
-                                         const framework::BlockDesc &block_desc,
-                                         const framework::Scope &scope,
-                                         bool test_mode) {
+template <typename TargetT>
+void DetectionOutOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   auto target_name = op_desc.Input("TargetBox").front();
   auto prior_box_name = op_desc.Input("PriorBox").front();
@@ -52,22 +47,28 @@ void DetectionOutOpConverter::operator()(const framework::proto::OpDesc &op,
         "Not support encode_center_size code_type in DetectionOut of anakin");
   }
 
-  engine_->AddOp(op_name, "DetectionOutput",
-                 {target_name, scores_name, prior_box_name}, {output_name});
-  engine_->AddOpAttr(op_name, "share_location", true);
-  engine_->AddOpAttr(op_name, "variance_encode_in_target", false);
-  engine_->AddOpAttr(op_name, "class_num", static_cast<int>(0));
-  engine_->AddOpAttr(op_name, "background_id", background_label);
-  engine_->AddOpAttr(op_name, "keep_top_k", keep_top_k);
-  engine_->AddOpAttr(op_name, "code_type", anakin_code_type);
-  engine_->AddOpAttr(op_name, "conf_thresh", score_threshold);
-  engine_->AddOpAttr(op_name, "nms_top_k", nms_top_k);
-  engine_->AddOpAttr(op_name, "nms_thresh", nms_threshold);
-  engine_->AddOpAttr(op_name, "nms_eta", nms_eta);
+  this->engine_->AddOp(op_name, "DetectionOutput",
+                       {target_name, scores_name, prior_box_name},
+                       {output_name});
+  this->engine_->AddOpAttr(op_name, "share_location", true);
+  this->engine_->AddOpAttr(op_name, "variance_encode_in_target", false);
+  this->engine_->AddOpAttr(op_name, "class_num", static_cast<int>(0));
+  this->engine_->AddOpAttr(op_name, "background_id", background_label);
+  this->engine_->AddOpAttr(op_name, "keep_top_k", keep_top_k);
+  this->engine_->AddOpAttr(op_name, "code_type", anakin_code_type);
+  this->engine_->AddOpAttr(op_name, "conf_thresh", score_threshold);
+  this->engine_->AddOpAttr(op_name, "nms_top_k", nms_top_k);
+  this->engine_->AddOpAttr(op_name, "nms_thresh", nms_threshold);
+  this->engine_->AddOpAttr(op_name, "nms_eta", nms_eta);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(detection_out, DetectionOutOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(detection_out,
+                                  DetectionOutOpConverter<::anakin::saber::NV>);
+#endif
+REGISTER_CPU_ANAKIN_OP_CONVERTER(detection_out,
+                                 DetectionOutOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/detection_out.h

@@ -22,7 +22,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class DetectionOutOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class DetectionOutOpConverter : public AnakinOpConverter<TargetT> {
  public:
   DetectionOutOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/dropout.cc

@@ -19,21 +19,16 @@
 
 using anakin::graph::GraphGlobalMem;
 using anakin::AK_FLOAT;
-using anakin::Precision;
-using anakin::saber::NV;
-using anakin::saber::X86;
 using anakin::saber::Shape;
-using anakin::PBlock;
-using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void DropoutOpConverter::operator()(const framework::proto::OpDesc &op,
-                                    const framework::BlockDesc &block_desc,
-                                    const framework::Scope &scope,
-                                    bool test_mode) {
+template <typename TargetT>
+void DropoutOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
   PADDLE_ENFORCE_EQ(op_desc.Output("Mask").size(), 1);
@@ -43,25 +38,30 @@ void DropoutOpConverter::operator()(const framework::proto::OpDesc &op,
   auto out_name = op_desc.Output("Out").front();
   auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
 
-  engine_->AddOp(op_name, "Scale", {x_name}, {out_name});
+  this->engine_->AddOp(op_name, "Scale", {x_name}, {out_name});
 
   auto dropout_prob = boost::get<float>(op_desc.GetAttr("dropout_prob"));
   auto factor = 1 - dropout_prob;
   Shape shape1(std::vector<int>({1, 1, 1, 1}));
   auto *weight1 =
-      GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(shape1);
+      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(shape1);
   auto *factor_data = static_cast<float *>(weight1->h_tensor().mutable_data());
   float weight1_data[] = {factor};
   std::copy(std::begin(weight1_data), std::end(weight1_data), factor_data);
 
-  engine_->AddOpAttr(op_name, "weight_1", *weight1);
-  engine_->AddOpAttr(op_name, "axis", 0);
-  engine_->AddOpAttr(op_name, "num_axes", 0);
-  engine_->AddOpAttr(op_name, "bias_term", false);
+  this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
+  this->engine_->AddOpAttr(op_name, "axis", 0);
+  this->engine_->AddOpAttr(op_name, "num_axes", 0);
+  this->engine_->AddOpAttr(op_name, "bias_term", false);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(dropout, DropoutOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(dropout,
+                                  DropoutOpConverter<::anakin::saber::NV>);
+#endif
+REGISTER_CPU_ANAKIN_OP_CONVERTER(dropout,
+                                 DropoutOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/dropout.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class DropoutOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class DropoutOpConverter : public AnakinOpConverter<TargetT> {
  public:
   DropoutOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/elementwise.cc

@@ -19,18 +19,15 @@
 
 using anakin::graph::GraphGlobalMem;
 using anakin::AK_FLOAT;
-using anakin::Precision;
-using anakin::saber::NV;
-using anakin::saber::X86;
 using anakin::saber::Shape;
-using anakin::PBlock;
 using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void ElementwiseAddOpConverter::operator()(
+template <typename TargetT>
+void ElementwiseAddOpConverter<TargetT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -43,14 +40,16 @@ void ElementwiseAddOpConverter::operator()(
   auto out_name = op_desc.Output("Out").front();
   auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
 
-  engine_->AddOp(op_name, "Eltwise", {x_name, y_name}, {out_name});
+  this->engine_->AddOp(op_name, "Eltwise", {x_name, y_name}, {out_name});
   std::string elementwise_type = "Add";
-  engine_->AddOpAttr<std::string>(op_name, "type", elementwise_type);
+  this->engine_->template AddOpAttr<std::string>(op_name, "type",
+                                                 elementwise_type);
   std::vector<float> coeff = {1.0, 1.0};
-  engine_->AddOpAttr<PTuple<float>>(op_name, "coeff", coeff);
+  this->engine_->template AddOpAttr<PTuple<float>>(op_name, "coeff", coeff);
 }
 
-void ElementwiseMulOpConverter::operator()(
+template <typename TargetT>
+void ElementwiseMulOpConverter<TargetT>::operator()(
     const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
     const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
@@ -63,26 +62,25 @@ void ElementwiseMulOpConverter::operator()(
   auto out_name = op_desc.Output("Out").front();
   auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
 
-  engine_->AddOp(op_name, "Scale", {x_name, y_name}, {out_name});
-  // Fill a number to weight_1 as a placeholder.
-  Shape shape1(std::vector<int>({1, 1, 1, 1}));
-  auto *weight1 =
-      GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(shape1);
-  auto *placeholder_data =
-      static_cast<float *>(weight1->h_tensor().mutable_data());
-  float weight1_data[] = {1};
-  std::copy(std::begin(weight1_data), std::end(weight1_data), placeholder_data);
-  engine_->AddOpAttr(op_name, "weight_1", *weight1);
-
-  auto axis = boost::get<int>(op_desc.GetAttr("axis"));
-  engine_->AddOpAttr(op_name, "axis", axis);
-  engine_->AddOpAttr(op_name, "num_axes", 1);
-  engine_->AddOpAttr(op_name, "bias_term", false);
+  this->engine_->AddOp(op_name, "Eltwise", {x_name, y_name}, {out_name});
+  std::string elementwise_type = "Prod";
+  this->engine_->template AddOpAttr<std::string>(op_name, "type",
+                                                 elementwise_type);
+  std::vector<float> coeff = {1.0, 1.0};
+  this->engine_->template AddOpAttr<PTuple<float>>(op_name, "coeff", coeff);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(elementwise_add, ElementwiseAddOpConverter);
-REGISTER_ANAKIN_OP_CONVERTER(elementwise_mul, ElementwiseMulOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(
+    elementwise_add, ElementwiseAddOpConverter<::anakin::saber::NV>);
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(
+    elementwise_mul, ElementwiseMulOpConverter<::anakin::saber::NV>);
+#endif
+REGISTER_CPU_ANAKIN_OP_CONVERTER(
+    elementwise_add, ElementwiseAddOpConverter<::anakin::saber::X86>);
+REGISTER_CPU_ANAKIN_OP_CONVERTER(
+    elementwise_mul, ElementwiseMulOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/elementwise.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class ElementwiseAddOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class ElementwiseAddOpConverter : public AnakinOpConverter<TargetT> {
  public:
   ElementwiseAddOpConverter() = default;
 
@@ -33,7 +34,8 @@ class ElementwiseAddOpConverter : public AnakinOpConverter {
  private:
 };
 
-class ElementwiseMulOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class ElementwiseMulOpConverter : public AnakinOpConverter<TargetT> {
  public:
   ElementwiseMulOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/fc.cc

@@ -19,17 +19,16 @@
 
 using anakin::graph::GraphGlobalMem;
 using anakin::AK_FLOAT;
-using anakin::saber::NV;
 using anakin::saber::Shape;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void FcBaseOpConverter::operator()(const framework::proto::OpDesc &op,
-                                   const framework::BlockDesc &block_desc,
-                                   const framework::Scope &scope,
-                                   bool test_mode) {
+template <typename TargetT>
+void FcBaseOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   auto input_names = op_desc.InputNames();
   bool with_bias = input_names.size() == 3;
@@ -51,13 +50,13 @@ void FcBaseOpConverter::operator()(const framework::proto::OpDesc &op,
   auto input_name = op_desc.Input(i_name).front();
   auto output_name = op_desc.Output("Out").front();
 
-  engine_->AddOp(op_name, "Dense", {input_name}, {output_name});
-  engine_->AddOpAttr(op_name, "bias_term", with_bias);
-  engine_->AddOpAttr(op_name, "axis", 1);
+  this->engine_->AddOp(op_name, "Dense", {input_name}, {output_name});
+  this->engine_->AddOpAttr(op_name, "bias_term", with_bias);
+  this->engine_->AddOpAttr(op_name, "axis", 1);
 
   auto weight_shape = framework::vectorize2int(y_t->dims());
   int out_dim = weight_shape[1];
-  engine_->AddOpAttr(op_name, "out_dim", out_dim);
+  this->engine_->AddOpAttr(op_name, "out_dim", out_dim);
   const int w_m = weight_shape[0];
   const int w_k = weight_shape[1];
 
@@ -79,12 +78,13 @@ void FcBaseOpConverter::operator()(const framework::proto::OpDesc &op,
     }
   }
   auto *weight1 =
-      GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(anakin_shape);
+      GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
+          anakin_shape);
   float *cpu_data = static_cast<float *>(weight1->h_tensor().mutable_data());
   std::copy_n(trans_weight_data.data(), weight_tensor.numel(), cpu_data);
   weight1->d_tensor().set_shape(anakin_shape);
   weight1->d_tensor().copy_from(weight1->h_tensor());
-  engine_->AddOpAttr(op_name, "weight_1", *weight1);
+  this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
 
   // get bias
   if (with_bias) {
@@ -104,13 +104,14 @@ void FcBaseOpConverter::operator()(const framework::proto::OpDesc &op,
     // bias_shape.push_back(1);
     Shape anakin_bias_shape(bias_shape);
 
-    auto *weight2 = GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(
-        anakin_bias_shape);
+    auto *weight2 =
+        GraphGlobalMem<TargetT>::Global().template new_block<AK_FLOAT>(
+            anakin_bias_shape);
     float *cpu_data2 = static_cast<float *>(weight2->h_tensor().mutable_data());
     std::copy_n(bias_data, bias_tensor.numel(), cpu_data2);
     weight2->d_tensor().set_shape(anakin_bias_shape);
     weight2->d_tensor().copy_from(weight2->h_tensor());
-    engine_->AddOpAttr(op_name, "weight_2", *weight2);
+    this->engine_->AddOpAttr(op_name, "weight_2", *weight2);
   }
 }
 
@@ -118,5 +119,10 @@ void FcBaseOpConverter::operator()(const framework::proto::OpDesc &op,
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(mul, MulOpConverter);
-REGISTER_ANAKIN_OP_CONVERTER(fc, FcOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(mul, MulOpConverter<::anakin::saber::NV>);
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(fc, FcOpConverter<::anakin::saber::NV>);
+#endif
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(mul, MulOpConverter<::anakin::saber::X86>);
+REGISTER_CPU_ANAKIN_OP_CONVERTER(fc, FcOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/fc.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class FcBaseOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class FcBaseOpConverter : public AnakinOpConverter<TargetT> {
  public:
   FcBaseOpConverter() = default;
 
@@ -32,13 +33,15 @@ class FcBaseOpConverter : public AnakinOpConverter {
 };
 
 // with bias
-class FcOpConverter : public FcBaseOpConverter {
+template <typename TargetT>
+class FcOpConverter : public FcBaseOpConverter<TargetT> {
  public:
   FcOpConverter() = default;
 };
 
 // without bias
-class MulOpConverter : public FcBaseOpConverter {
+template <typename TargetT>
+class MulOpConverter : public FcBaseOpConverter<TargetT> {
  public:
   MulOpConverter() = default;
 };

paddle/fluid/inference/anakin/convert/flatten.cc

@@ -15,20 +15,16 @@
 #include "paddle/fluid/inference/anakin/convert/flatten.h"
 #include <vector>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::saber::NV;
-using anakin::saber::Shape;
 using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void FlattenOpConverter::operator()(const framework::proto::OpDesc &op,
-                                    const framework::BlockDesc &block_desc,
-                                    const framework::Scope &scope,
-                                    bool test_mode) {
+template <typename TargetT>
+void FlattenOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1UL);
   PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1UL);
@@ -41,12 +37,17 @@ void FlattenOpConverter::operator()(const framework::proto::OpDesc &op,
 
   std::vector<int> out_dims = {0, -1, 1, 1};
   auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
-  engine_->AddOp(op_name, "Reshape", {input}, {output});
-  engine_->AddOpAttr<PTuple<int>>(op_name, "dims", out_dims);
+  this->engine_->AddOp(op_name, "Reshape", {input}, {output});
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "dims", out_dims);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(flatten, FlattenOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(flatten,
+                                  FlattenOpConverter<::anakin::saber::NV>);
+#endif
+REGISTER_CPU_ANAKIN_OP_CONVERTER(flatten,
+                                 FlattenOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/flatten.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class FlattenOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class FlattenOpConverter : public AnakinOpConverter<TargetT> {
  public:
   FlattenOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/im2sequence.cc

@@ -17,23 +17,16 @@
 #include <string>
 #include <vector>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::Precision;
-using anakin::saber::NV;
-using anakin::saber::X86;
-using anakin::saber::Shape;
-using anakin::PBlock;
 using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void Im2SequenceConverter::operator()(const framework::proto::OpDesc &op,
-                                      const framework::BlockDesc &block_desc,
-                                      const framework::Scope &scope,
-                                      bool test_mode) {
+template <typename TargetT>
+void Im2SequenceConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
   PADDLE_ENFORCE_EQ(op_desc.Output("Y").size(), 0);
@@ -43,21 +36,24 @@ void Im2SequenceConverter::operator()(const framework::proto::OpDesc &op,
   auto out_name = op_desc.Output("Out").front();
   auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
 
-  engine_->AddOp(op_name, "Im2Sequence", {x_name}, {out_name});
+  this->engine_->AddOp(op_name, "Im2Sequence", {x_name}, {out_name});
 
   std::vector<int> dilations = {1, 1};
   auto paddings = boost::get<std::vector<int>>(op_desc.GetAttr("paddings"));
   auto strides = boost::get<std::vector<int>>(op_desc.GetAttr("strides"));
   auto kernels = boost::get<std::vector<int>>(op_desc.GetAttr("kernels"));
 
-  engine_->AddOpAttr<PTuple<int>>(op_name, "paddings", paddings);
-  engine_->AddOpAttr<PTuple<int>>(op_name, "strides", strides);
-  engine_->AddOpAttr<PTuple<int>>(op_name, "window_size", kernels);
-  engine_->AddOpAttr<PTuple<int>>(op_name, "dilations", dilations);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "paddings", paddings);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "strides", strides);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "window_size",
+                                                 kernels);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "dilations",
+                                                 dilations);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(im2sequence, Im2SequenceConverter);
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(im2sequence,
+                                  Im2SequenceConverter<::anakin::saber::NV>);

paddle/fluid/inference/anakin/convert/im2sequence.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class Im2SequenceConverter : public AnakinOpConverter {
+template <typename TargetT>
+class Im2SequenceConverter : public AnakinOpConverter<TargetT> {
  public:
   Im2SequenceConverter() = default;
 

paddle/fluid/inference/anakin/convert/op_converter.h

@@ -32,10 +32,10 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-using AnakinNvEngine =
-    AnakinEngine<::anakin::saber::NV, ::anakin::Precision::FP32>;
-
+template <typename TargetT>
 class AnakinOpConverter {
+  using AnakinEngineT = AnakinEngine<TargetT, ::anakin::Precision::FP32>;
+
  public:
   AnakinOpConverter() = default;
 
@@ -45,7 +45,7 @@ class AnakinOpConverter {
   void ConvertOp(const framework::proto::OpDesc &op,
                  const framework::BlockDesc &block_desc,
                  const std::unordered_set<std::string> &parameters,
-                 const framework::Scope &scope, AnakinNvEngine *engine,
+                 const framework::Scope &scope, AnakinEngineT *engine,
                  bool test_mode = false) {
     framework::OpDesc op_desc(op, nullptr);
     std::string op_type = op_desc.Type();
@@ -65,7 +65,7 @@ class AnakinOpConverter {
 
   void ConvertBlock(framework::BlockDesc *block_desc,
                     const std::unordered_set<std::string> &parameters,
-                    const framework::Scope &scope, AnakinNvEngine *engine) {
+                    const framework::Scope &scope, AnakinEngineT *engine) {
     std::unique_lock<std::mutex> lock(mutex_);
     framework::proto::BlockDesc *block = block_desc->Proto();
     for (auto i = 0; i < block->ops_size(); i++) {
@@ -79,7 +79,7 @@ class AnakinOpConverter {
       framework::BlockDesc *block_desc, framework::Scope *scope,
       const std::vector<std::string> &inputs,
       const std::unordered_set<std::string> &parameters,
-      const std::vector<std::string> &outputs, AnakinNvEngine *engine) {
+      const std::vector<std::string> &outputs, AnakinEngineT *engine) {
     ConvertBlock(block_desc, parameters, *scope, engine);
     // if the max_batch size
     int max_batch_size = engine->GetMaxBatchSize();
@@ -128,40 +128,60 @@ class AnakinOpConverter {
     engine->InitNet();
   }
 
-  void SetEngine(AnakinNvEngine *engine) { engine_ = engine; }
+  void SetEngine(AnakinEngineT *engine) { engine_ = engine; }
   virtual ~AnakinOpConverter() {}
 
  protected:
   bool test_mode_;
-  AnakinNvEngine *engine_{nullptr};
+  AnakinEngineT *engine_{nullptr};
 
  private:
-  std::unordered_map<std::string, AnakinOpConverter *> converters_;
+  std::unordered_map<std::string, AnakinOpConverter<TargetT> *> converters_;
   framework::Scope *scope_{nullptr};
   std::mutex mutex_;
 };
 
+template class AnakinOpConverter<::anakin::saber::NV>;
+template class AnakinOpConverter<::anakin::saber::X86>;
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-#define REGISTER_ANAKIN_OP_CONVERTER(op_type__, Converter__)               \
-  struct anakin_##op_type__##_converter                                    \
-      : public ::paddle::framework::Registrar {                            \
-    anakin_##op_type__##_converter() {                                     \
-      LOG(INFO) << "register convert " << #op_type__;                      \
-      ::paddle::inference::Registry<                                       \
-          ::paddle::inference::anakin::AnakinOpConverter>::Global()        \
-          .Register<::paddle::inference::anakin::Converter__>(#op_type__); \
-    }                                                                      \
-  };                                                                       \
-  anakin_##op_type__##_converter anakin_##op_type__##_converter__;         \
-  int TouchConverterRegister_anakin_##op_type__() {                        \
-    anakin_##op_type__##_converter__.Touch();                              \
-    return 0;                                                              \
+#define REGISTER_ANAKIN_OP_CONVERTER_BASE(op_type__, Converter__,              \
+                                          place_type__, place_class__)         \
+  struct anakin_##op_type__##_##place_type__##_converter                       \
+      : public ::paddle::framework::Registrar {                                \
+    anakin_##op_type__##_##place_type__##_converter() {                        \
+      LOG(INFO) << "register convert " << #op_type__ << " ";                   \
+      ::paddle::inference::Registry<                                           \
+          ::paddle::inference::anakin::AnakinOpConverter<place_class__>>::     \
+          Global()                                                             \
+              .Register<::paddle::inference::anakin::Converter__>(#op_type__); \
+    }                                                                          \
+  };                                                                           \
+  anakin_##op_type__##_##place_type__##_converter                              \
+      anakin_##op_type__##_##place_type__##_converter__;                       \
+  int TouchConverterRegister_anakin_##op_type__##_##place_type__() {           \
+    anakin_##op_type__##_##place_type__##_converter__.Touch();                 \
+    return 0;                                                                  \
   }
 
-#define USE_ANAKIN_CONVERTER(op_type__)                             \
-  extern int TouchConverterRegister_anakin_##op_type__();           \
-  int use_op_converter_anakin_##op_type__ __attribute__((unused)) = \
-      TouchConverterRegister_anakin_##op_type__();
+#define REGISTER_CUDA_ANAKIN_OP_CONVERTER(op_type__, Converter__) \
+  REGISTER_ANAKIN_OP_CONVERTER_BASE(op_type__, Converter__, CUDA, \
+                                    ::anakin::saber::NV)
+
+#define REGISTER_CPU_ANAKIN_OP_CONVERTER(op_type__, Converter__) \
+  REGISTER_ANAKIN_OP_CONVERTER_BASE(op_type__, Converter__, CPU, \
+                                    ::anakin::saber::X86)
+
+#define USE_ANAKIN_CONVERTER_BASE(op_type__, place_type__)                 \
+  extern int TouchConverterRegister_anakin_##op_type__##_##place_type__(); \
+  int use_op_converter_anakin_##op_type__##_##place_type__                 \
+      __attribute__((unused)) =                                            \
+          TouchConverterRegister_anakin_##op_type__##_##place_type__();
+
+#define USE_ANAKIN_CONVERTER(op_type__) \
+  USE_ANAKIN_CONVERTER_BASE(op_type__, CUDA)
+
+#define USE_CPU_ANAKIN_CONVERTER(op_type__) \
+  USE_ANAKIN_CONVERTER_BASE(op_type__, CPU)

paddle/fluid/inference/anakin/convert/pool2d.cc

@@ -17,23 +17,16 @@
 #include <string>
 #include <vector>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::Precision;
-using anakin::saber::NV;
-using anakin::saber::X86;
-using anakin::saber::Shape;
-using anakin::PBlock;
 using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void Pool2dOpConverter::operator()(const framework::proto::OpDesc &op,
-                                   const framework::BlockDesc &block_desc,
-                                   const framework::Scope &scope,
-                                   bool test_mode) {
+template <typename TargetT>
+void Pool2dOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
   PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
@@ -65,17 +58,22 @@ void Pool2dOpConverter::operator()(const framework::proto::OpDesc &op,
     PADDLE_THROW("TensorRT unsupported pooling type!");
   }
 
-  engine_->AddOp(op_name, "Pooling", {x_name}, {y_name});
-  engine_->AddOpAttr<PTuple<int>>(op_name, "pool_size", ksize);
-  engine_->AddOpAttr<PTuple<int>>(op_name, "strides", strides);
-  engine_->AddOpAttr<PTuple<int>>(op_name, "padding", paddings);
-  engine_->AddOpAttr(op_name, "method", anakin_pool_type);
-  engine_->AddOpAttr(op_name, "global_pooling", global_pooling);
-  engine_->AddOpAttr(op_name, "cmp_out_shape_floor_as_conv", !ceil_mode);
+  this->engine_->AddOp(op_name, "Pooling", {x_name}, {y_name});
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "pool_size", ksize);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "strides", strides);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "padding", paddings);
+  this->engine_->AddOpAttr(op_name, "method", anakin_pool_type);
+  this->engine_->AddOpAttr(op_name, "global_pooling", global_pooling);
+  this->engine_->AddOpAttr(op_name, "cmp_out_shape_floor_as_conv", !ceil_mode);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(pool2d, Pool2dOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(pool2d,
+                                  Pool2dOpConverter<::anakin::saber::NV>);
+#endif
+REGISTER_CPU_ANAKIN_OP_CONVERTER(pool2d,
+                                 Pool2dOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/pool2d.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class Pool2dOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class Pool2dOpConverter : public AnakinOpConverter<TargetT> {
  public:
   Pool2dOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/relu.cc

@@ -16,19 +16,14 @@
 #include <algorithm>
 #include <map>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::saber::NV;
-using anakin::saber::Shape;
-
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void ReluOpConverter::operator()(const framework::proto::OpDesc &op,
-                                 const framework::BlockDesc &block_desc,
-                                 const framework::Scope &scope,
-                                 bool test_mode) {
+template <typename TargetT>
+void ReluOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
   PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
@@ -37,14 +32,14 @@ void ReluOpConverter::operator()(const framework::proto::OpDesc &op,
   auto input_name = op_desc.Input("X").front();
   auto output_name = op_desc.Output("Out").front();
 
-  engine_->AddOp(op_name, "ReLU", {input_name}, {output_name});
-  engine_->AddOpAttr(op_name, "alpha", 0);
+  this->engine_->AddOp(op_name, "ReLU", {input_name}, {output_name});
+  this->engine_->AddOpAttr(op_name, "alpha", 0);
 }
 
-void LeakyReluOpConverter::operator()(const framework::proto::OpDesc &op,
-                                      const framework::BlockDesc &block_desc,
-                                      const framework::Scope &scope,
-                                      bool test_mode) {
+template <typename TargetT>
+void LeakyReluOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
   PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
@@ -54,13 +49,19 @@ void LeakyReluOpConverter::operator()(const framework::proto::OpDesc &op,
   auto output_name = op_desc.Output("Out").front();
 
   float alpha = boost::get<float>(op_desc.GetAttr("alpha"));
-  engine_->AddOp(op_name, "ReLU", {input_name}, {output_name});
-  engine_->AddOpAttr(op_name, "alpha", alpha);
+  this->engine_->AddOp(op_name, "ReLU", {input_name}, {output_name});
+  this->engine_->AddOpAttr(op_name, "alpha", alpha);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(relu, ReluOpConverter);
-REGISTER_ANAKIN_OP_CONVERTER(leaky_relu, LeakyReluOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(relu, ReluOpConverter<::anakin::saber::NV>);
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(leaky_relu,
+                                  LeakyReluOpConverter<::anakin::saber::NV>);
+#endif
+REGISTER_CPU_ANAKIN_OP_CONVERTER(relu, ReluOpConverter<::anakin::saber::X86>);
+REGISTER_CPU_ANAKIN_OP_CONVERTER(leaky_relu,
+                                 LeakyReluOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/relu.h

@@ -22,7 +22,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class ReluOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class ReluOpConverter : public AnakinOpConverter<TargetT> {
  public:
   ReluOpConverter() = default;
 
@@ -33,7 +34,8 @@ class ReluOpConverter : public AnakinOpConverter {
   virtual ~ReluOpConverter() {}
 };
 
-class LeakyReluOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class LeakyReluOpConverter : public AnakinOpConverter<TargetT> {
  public:
   LeakyReluOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/reshape.cc

@@ -15,20 +15,16 @@
 #include "paddle/fluid/inference/anakin/convert/reshape.h"
 #include <vector>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::saber::NV;
-using anakin::saber::Shape;
 using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void ReshapeOpConverter::operator()(const framework::proto::OpDesc &op,
-                                    const framework::BlockDesc &block_desc,
-                                    const framework::Scope &scope,
-                                    bool test_mode) {
+template <typename TargetT>
+void ReshapeOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1UL);
   PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1UL);
@@ -37,17 +33,23 @@ void ReshapeOpConverter::operator()(const framework::proto::OpDesc &op,
   auto output = op_desc.Output("Out").front();
 
   auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
-  engine_->AddOp(op_name, "Reshape", {input}, {output});
+  this->engine_->AddOp(op_name, "Reshape", {input}, {output});
 
   auto shape = boost::get<std::vector<int>>(op_desc.GetAttr("shape"));
   if (shape.size() < 4) {
     shape.insert(shape.end(), 4 - shape.size(), 1);
   }
-  engine_->AddOpAttr<PTuple<int>>(op_name, "dims", shape);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "dims", shape);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(reshape, ReshapeOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(reshape,
+                                  ReshapeOpConverter<::anakin::saber::NV>);
+#endif
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(reshape,
+                                 ReshapeOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/reshape.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class ReshapeOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class ReshapeOpConverter : public AnakinOpConverter<TargetT> {
  public:
   ReshapeOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/roi_align.cc

@@ -25,10 +25,10 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-void RoiAlignOpConverter::operator()(const framework::proto::OpDesc &op,
-                                     const framework::BlockDesc &block_desc,
-                                     const framework::Scope &scope,
-                                     bool test_mode) {
+template <typename TargetT>
+void RoiAlignOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
   PADDLE_ENFORCE_EQ(op_desc.Input("ROIs").size(), 1);
@@ -44,16 +44,21 @@ void RoiAlignOpConverter::operator()(const framework::proto::OpDesc &op,
   auto pooled_width = boost::get<int>(op_desc.GetAttr("pooled_width"));
   auto sampling_ratio = boost::get<int>(op_desc.GetAttr("sampling_ratio"));
 
-  engine_->AddOp(op_name, "RoiAlign", {input_x_name, input_rois_name},
-                 {output_name});
-  engine_->AddOpAttr(op_name, "spatial_scale", spatial_scale);
-  engine_->AddOpAttr(op_name, "pooled_height", pooled_height);
-  engine_->AddOpAttr(op_name, "pooled_width", pooled_width);
-  engine_->AddOpAttr(op_name, "sampling_ratio", sampling_ratio);
+  this->engine_->AddOp(op_name, "RoiAlign", {input_x_name, input_rois_name},
+                       {output_name});
+  this->engine_->AddOpAttr(op_name, "spatial_scale", spatial_scale);
+  this->engine_->AddOpAttr(op_name, "pooled_height", pooled_height);
+  this->engine_->AddOpAttr(op_name, "pooled_width", pooled_width);
+  this->engine_->AddOpAttr(op_name, "sampling_ratio", sampling_ratio);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(roi_align, RoiAlignOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(roi_align,
+                                  RoiAlignOpConverter<::anakin::saber::NV>);
+#endif
+REGISTER_CPU_ANAKIN_OP_CONVERTER(roi_align,
+                                 RoiAlignOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/roi_align.h

@@ -22,7 +22,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class RoiAlignOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class RoiAlignOpConverter : public AnakinOpConverter<TargetT> {
  public:
   RoiAlignOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/scale.cc

@@ -16,19 +16,14 @@
 #include <algorithm>
 #include <map>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::saber::NV;
-using anakin::saber::Shape;
-
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void ScaleOpConverter::operator()(const framework::proto::OpDesc &op,
-                                  const framework::BlockDesc &block_desc,
-                                  const framework::Scope &scope,
-                                  bool test_mode) {
+template <typename TargetT>
+void ScaleOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
   PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
@@ -44,14 +39,14 @@ void ScaleOpConverter::operator()(const framework::proto::OpDesc &op,
   PADDLE_ENFORCE(bias_after_scale,
                  "The anakin scale layer only support bias after scale now.");
 
-  engine_->AddOp(op_name, "Power", {input_name}, {output_name});
-  engine_->AddOpAttr(op_name, "shift", bias);
-  engine_->AddOpAttr(op_name, "scale", scale);
-  engine_->AddOpAttr(op_name, "power", static_cast<float>(1.0));
+  this->engine_->AddOp(op_name, "Power", {input_name}, {output_name});
+  this->engine_->AddOpAttr(op_name, "shift", bias);
+  this->engine_->AddOpAttr(op_name, "scale", scale);
+  this->engine_->AddOpAttr(op_name, "power", static_cast<float>(1.0));
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(scale, ScaleOpConverter);
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(scale, ScaleOpConverter<::anakin::saber::NV>);

paddle/fluid/inference/anakin/convert/scale.h

@@ -22,7 +22,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class ScaleOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class ScaleOpConverter : public AnakinOpConverter<TargetT> {
  public:
   ScaleOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/softmax.cc

@@ -14,19 +14,14 @@
 
 #include "paddle/fluid/inference/anakin/convert/softmax.h"
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::saber::NV;
-using anakin::saber::Shape;
-
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void SoftMaxOpConverter::operator()(const framework::proto::OpDesc &op,
-                                    const framework::BlockDesc &block_desc,
-                                    const framework::Scope &scope,
-                                    bool test_mode) {
+template <typename TargetT>
+void SoftMaxOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1UL);
 
@@ -41,12 +36,18 @@ void SoftMaxOpConverter::operator()(const framework::proto::OpDesc &op,
   auto input_shape_in_fluid = input_var_desc->GetShape();
   size_t input_dims = input_shape_in_fluid.size();
 
-  engine_->AddOp(op_name, "Softmax", {input}, {output});
-  engine_->AddOpAttr(op_name, "axis", static_cast<int>(input_dims - 1));
+  this->engine_->AddOp(op_name, "Softmax", {input}, {output});
+  this->engine_->AddOpAttr(op_name, "axis", static_cast<int>(input_dims - 1));
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(softmax, SoftMaxOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(softmax,
+                                  SoftMaxOpConverter<::anakin::saber::NV>);
+#endif
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(softmax,
+                                 SoftMaxOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/softmax.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class SoftMaxOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class SoftMaxOpConverter : public AnakinOpConverter<TargetT> {
  public:
   SoftMaxOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/split.cc

@@ -16,23 +16,16 @@
 #include <algorithm>
 #include <vector>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::Precision;
-using anakin::saber::NV;
-using anakin::saber::X86;
-using anakin::saber::Shape;
-using anakin::PBlock;
 using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void SplitOpConverter::operator()(const framework::proto::OpDesc &op,
-                                  const framework::BlockDesc &block_desc,
-                                  const framework::Scope &scope,
-                                  bool test_mode) {
+template <typename TargetT>
+void SplitOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   auto input_name = op_desc.Input("X").front();
   auto y_names = op_desc.Output("Out");
@@ -51,14 +44,19 @@ void SplitOpConverter::operator()(const framework::proto::OpDesc &op,
     num_sum += output_lengths[i];
     slice_point.push_back(num_sum);
   }
-  engine_->AddOp(op_name, "Slice", {input_name}, y_names);
-  engine_->AddOpAttr(op_name, "axis", axis);
-  engine_->AddOpAttr<PTuple<int>>(op_name, "slice_point", slice_point);
+  this->engine_->AddOp(op_name, "Slice", {input_name}, y_names);
+  this->engine_->AddOpAttr(op_name, "axis", axis);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "slice_point",
+                                                 slice_point);
   // slice_dim is useless in anakin
-  engine_->AddOpAttr(op_name, "slice_dim", 4);
+  this->engine_->AddOpAttr(op_name, "slice_dim", 4);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
-REGISTER_ANAKIN_OP_CONVERTER(split, SplitOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(split, SplitOpConverter<::anakin::saber::NV>);
+#endif
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(split, SplitOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/split.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class SplitOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class SplitOpConverter : public AnakinOpConverter<TargetT> {
  public:
   SplitOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/sum.cc

@@ -17,22 +17,17 @@
 #include <string>
 #include <vector>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::Precision;
-using anakin::saber::NV;
-using anakin::saber::X86;
-using anakin::saber::Shape;
-using anakin::PBlock;
 using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void SumOpConverter::operator()(const framework::proto::OpDesc &op,
-                                const framework::BlockDesc &block_desc,
-                                const framework::Scope &scope, bool test_mode) {
+template <typename TargetT>
+void SumOpConverter<TargetT>::operator()(const framework::proto::OpDesc &op,
+                                         const framework::BlockDesc &block_desc,
+                                         const framework::Scope &scope,
+                                         bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 2);
   PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
@@ -43,13 +38,17 @@ void SumOpConverter::operator()(const framework::proto::OpDesc &op,
 
   std::vector<float> coeff = {1, 1};
   std::string elementwise_type = "Add";
-  engine_->AddOp(op_name, "Eltwise", input_names, {out_name});
-  engine_->AddOpAttr<PTuple<float>>(op_name, "coeff", coeff);
-  engine_->AddOpAttr<std::string>(op_name, "type", elementwise_type);
+  this->engine_->AddOp(op_name, "Eltwise", input_names, {out_name});
+  this->engine_->template AddOpAttr<PTuple<float>>(op_name, "coeff", coeff);
+  this->engine_->template AddOpAttr<std::string>(op_name, "type",
+                                                 elementwise_type);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(sum, SumOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(sum, SumOpConverter<::anakin::saber::NV>);
+#endif
+REGISTER_CPU_ANAKIN_OP_CONVERTER(sum, SumOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/sum.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class SumOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class SumOpConverter : public AnakinOpConverter<TargetT> {
  public:
   SumOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/test_activation_op.cc

@@ -21,12 +21,14 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-static void test_activation_op(const std::string &op_type) {
-  auto *converter = Registry<AnakinOpConverter>::Global().Lookup(op_type);
-  PADDLE_ENFORCE(converter != nullptr);
+template <typename TargetT>
+static void test_activation_op(const std::string& op_type,
+                               const platform::DeviceContext& context,
+                               bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
   validator.DeclInputVar("act-X", {10, 6, 1, 1});
   validator.DeclOutputVar("act-Out", {10, 6, 1, 1});
   framework::OpDesc desc;
@@ -41,13 +43,42 @@ static void test_activation_op(const std::string &op_type) {
   validator.Execute(5);
 }
 
-TEST(sigm_op, test) { test_activation_op("sigmoid"); }
-TEST(tanh_op, test) { test_activation_op("tanh"); }
+#ifdef PADDLE_WITH_CUDA
+TEST(sigm_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_activation_op<::anakin::saber::NV>("sigmoid", ctx, true);
+}
+
+TEST(tanh_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_activation_op<::anakin::saber::NV>("tanh", ctx, true);
+}
+#endif
+
+TEST(sigm_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_activation_op<::anakin::saber::X86>("sigmoid", ctx, false);
+}
+
+TEST(tanh_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_activation_op<::anakin::saber::X86>("tanh", ctx, false);
+}
+
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
 USE_OP(sigmoid);
 USE_OP(tanh);
+
+USE_CPU_ANAKIN_CONVERTER(sigmoid);
+USE_CPU_ANAKIN_CONVERTER(tanh);
+#ifdef PADDLE_WITH_CUDA
 USE_ANAKIN_CONVERTER(sigmoid);
 USE_ANAKIN_CONVERTER(tanh);
+#endif

paddle/fluid/inference/anakin/convert/test_affine_channel_op.cc

@@ -21,16 +21,19 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-TEST(affine_channel, native) {
+template <typename TargetT>
+void test_affine_channel_op(const platform::DeviceContext& context,
+                            bool use_gpu) {
   // Declare the difference between the inputs.
   std::unordered_set<std::string> parameters({"scale", "bias"});
 
   framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
   validator.DeclInputVar("x", {1, 3, 5, 2});
   validator.DeclOutputVar("out", {1, 3, 5, 2});
-  validator.DeclParamVar("scale", {1, 3, 1, 1});
-  validator.DeclParamVar("bias", {1, 3, 1, 1});
+  validator.DeclParamVar("scale", {3});
+  validator.DeclParamVar("bias", {3});
 
   // Prepare Op descriptions.
   framework::OpDesc desc;
@@ -47,9 +50,26 @@ TEST(affine_channel, native) {
   validator.Execute(1);
 }
 
+#ifdef PADDLE_WITH_CUDA
+TEST(affine_channel_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_affine_channel_op<::anakin::saber::NV>(ctx, true);
+}
+#endif
+
+TEST(affine_channel_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_affine_channel_op<::anakin::saber::X86>(ctx, false);
+}
+
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
 USE_OP(affine_channel);
+USE_CPU_ANAKIN_CONVERTER(affine_channel);
+#ifdef PADDLE_WITH_CUDA
 USE_ANAKIN_CONVERTER(affine_channel);
+#endif

paddle/fluid/inference/anakin/convert/test_batch_norm_op.cc

@@ -19,12 +19,14 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-TEST(batch_norm_op, test) {
+template <typename TargetT>
+void test_batchnorm_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters(
       {"batch_norm_scale", "batch_norm_bias", "batch_norm_mean",
        "batch_norm_variance"});
   framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
   std::vector<int> param_shape{2};
 
   validator.DeclInputVar("batch_norm_X", {1, 2, 5, 5});
@@ -64,8 +66,26 @@ TEST(batch_norm_op, test) {
   validator.Execute(1, neglected_output);
 }
 
+#ifdef PADDLE_WITH_CUDA
+TEST(batch_norm_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_batchnorm_op<::anakin::saber::NV>(ctx, true);
+}
+#endif
+
+TEST(batch_norm_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_batchnorm_op<::anakin::saber::X86>(ctx, false);
+}
+
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 USE_OP(batch_norm);
+USE_CPU_ANAKIN_CONVERTER(batch_norm);
+
+#ifdef PADDLE_WITH_CUDA
 USE_ANAKIN_CONVERTER(batch_norm);
+#endif

paddle/fluid/inference/anakin/convert/test_concat_op.cc

@@ -21,10 +21,12 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-TEST(concat_op, test) {
+template <typename TargetT>
+void test_concat_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters({""});
   framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
   validator.DeclInputVar("concat_x1", {1, 2, 1, 1});
   validator.DeclInputVar("concat_x2", {1, 3, 1, 1});
   validator.DeclInputVar("concat_x3", {1, 1, 1, 1});
@@ -44,31 +46,26 @@ TEST(concat_op, test) {
   validator.Execute(1);
 }
 
-TEST(concat_op, test2) {
-  std::unordered_set<std::string> parameters({""});
-  framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
-  validator.DeclInputVar("concat_x1", {1, 4});
-  validator.DeclInputVar("concat_x2", {3, 4});
-  validator.DeclInputVar("concat_x3", {2, 4});
-  validator.DeclOutputVar("concat_out", {6, 4});
-
-  // Prepare Op description
-  framework::OpDesc desc;
-  desc.SetType("concat");
-  desc.SetInput("X", {"concat_x1", "concat_x2", "concat_x3"});
-  desc.SetOutput("Out", {"concat_out"});
-
-  int axis = 0;
-  desc.SetAttr("axis", axis);
-
-  validator.SetOp(*desc.Proto());
+#ifdef PADDLE_WITH_CUDA
+TEST(concat_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_concat_op<::anakin::saber::NV>(ctx, true);
+}
+#endif
 
-  validator.Execute(1);
+TEST(concat_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_concat_op<::anakin::saber::X86>(ctx, false);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 USE_OP(concat);
+USE_CPU_ANAKIN_CONVERTER(concat);
+
+#ifdef PADDLE_WITH_CUDA
 USE_ANAKIN_CONVERTER(concat);
+#endif

paddle/fluid/inference/anakin/convert/test_conv2d_op.cc

@@ -21,13 +21,12 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-TEST(conv2d_op, test) {
-  auto* conv2d_converter =
-      Registry<AnakinOpConverter>::Global().Lookup("conv2d");
-  ASSERT_TRUE(conv2d_converter != nullptr);
+template <typename TargetT>
+void test_conv2d_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters({"conv2d-Y"});
   framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
   validator.DeclInputVar("conv2d-X", {1, 3, 3, 3});
   validator.DeclParamVar("conv2d-Y", {4, 3, 1, 1});
   validator.DeclOutputVar("conv2d-Out", {1, 4, 3, 3});
@@ -54,9 +53,27 @@ TEST(conv2d_op, test) {
   validator.Execute(3);
 }
 
+#ifdef PADDLE_WITH_CUDA
+TEST(conv2d_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_conv2d_op<::anakin::saber::NV>(ctx, true);
+}
+#endif
+
+TEST(conv2d_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_conv2d_op<::anakin::saber::X86>(ctx, false);
+}
+
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
 USE_OP(conv2d);
+USE_CPU_ANAKIN_CONVERTER(conv2d);
+
+#ifdef PADDLE_WITH_CUDA
 USE_ANAKIN_CONVERTER(conv2d);
+#endif

paddle/fluid/inference/anakin/convert/test_dropout_op.cc

@@ -21,10 +21,12 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-TEST(dropout_op, native) {
+template <typename TargetT>
+void test_dropout_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
   validator.DeclInputVar("x", {1, 1, 2, 2});
   validator.DeclOutputVar("out", {1, 1, 2, 2});
   validator.DeclOutputVar("mask", {1, 1, 2, 2});
@@ -45,9 +47,26 @@ TEST(dropout_op, native) {
   validator.Execute(1, neglected_output);
 }
 
+#ifdef PADDLE_WITH_CUDA
+TEST(dropout_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_dropout_op<::anakin::saber::NV>(ctx, true);
+}
+#endif
+
+TEST(dropout_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_dropout_op<::anakin::saber::X86>(ctx, false);
+}
+
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
 USE_OP(dropout);
+USE_CPU_ANAKIN_CONVERTER(dropout);
+#ifdef PADDLE_WITH_CUDA
 USE_ANAKIN_CONVERTER(dropout);
+#endif

paddle/fluid/inference/anakin/convert/test_elementwise_op.cc

@@ -21,10 +21,14 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-static void test_elementwise_op(const std::string &op_type) {
+template <typename TargetT>
+static void test_elementwise_op(const std::string& op_type,
+                                const platform::DeviceContext& context,
+                                bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
   validator.DeclInputVar("x", {1, 1, 2, 2});
   validator.DeclInputVar("y", {1, 1, 2, 2});
   validator.DeclOutputVar("out", {1, 1, 2, 2});
@@ -43,14 +47,41 @@ static void test_elementwise_op(const std::string &op_type) {
   validator.Execute(1);
 }
 
-TEST(elementwise_op, native_add) { test_elementwise_op("elementwise_add"); }
-TEST(elementwise_op, native_mul) { test_elementwise_op("elementwise_mul"); }
+#ifdef PADDLE_WITH_CUDA
+TEST(elementwise_op, native_add_gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_elementwise_op<::anakin::saber::NV>("elementwise_add", ctx, true);
+}
+TEST(elementwise_op, native_mul_gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_elementwise_op<::anakin::saber::NV>("elementwise_mul", ctx, true);
+}
+#endif
+
+TEST(elementwise_op, native_add_cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_elementwise_op<::anakin::saber::X86>("elementwise_add", ctx, false);
+}
+
+TEST(elementwise_op, native_mul_cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_elementwise_op<::anakin::saber::X86>("elementwise_mul", ctx, false);
+}
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
 USE_OP(elementwise_add);
-USE_ANAKIN_CONVERTER(elementwise_add);
 USE_OP(elementwise_mul);
+#ifdef PADDLE_WITH_CUDA
+USE_ANAKIN_CONVERTER(elementwise_add);
 USE_ANAKIN_CONVERTER(elementwise_mul);
+#endif
+
+USE_CPU_ANAKIN_CONVERTER(elementwise_add);
+USE_CPU_ANAKIN_CONVERTER(elementwise_mul);

paddle/fluid/inference/anakin/convert/test_fc_op.cc

@@ -20,13 +20,13 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-TEST(fc_op, test) {
-  auto* fc_converter = Registry<AnakinOpConverter>::Global().Lookup("fc");
-  ASSERT_TRUE(fc_converter);
-
+template <typename TargetT>
+void test_mul_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters({"mul_y"});
   framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
+
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
   validator.DeclInputVar("mul_x", {1, 1, 2, 2});
   validator.DeclParamVar("mul_y", {4, 2});
   validator.DeclOutputVar("mul_out", {1, 2});
@@ -42,9 +42,26 @@ TEST(fc_op, test) {
   validator.Execute(10);
 }
 
+#ifdef PADDLE_WITH_CUDA
+TEST(mul_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_mul_op<::anakin::saber::NV>(ctx, true);
+}
+#endif
+
+TEST(mul_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_mul_op<::anakin::saber::X86>(ctx, false);
+}
+
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
 USE_OP(mul);
+USE_CPU_ANAKIN_CONVERTER(fc);
+#ifdef PADDLE_WITH_CUDA
 USE_ANAKIN_CONVERTER(fc);
+#endif

paddle/fluid/inference/anakin/convert/test_flatten_op.cc

@@ -20,13 +20,12 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-TEST(flatten_op, test) {
-  auto *converter = Registry<AnakinOpConverter>::Global().Lookup("flatten");
-  ASSERT_TRUE(converter);
-
+template <typename TargetT>
+void test_flatten_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
   validator.DeclInputVar("flatten-X", {3, 10, 10, 4});
   validator.DeclOutputVar("flatten-Out", {3, 400, 1, 1});
   framework::OpDesc desc;
@@ -42,10 +41,27 @@ TEST(flatten_op, test) {
   validator.Execute(5);
 }
 
+#ifdef PADDLE_WITH_CUDA
+TEST(flatten_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_flatten_op<::anakin::saber::NV>(ctx, true);
+}
+#endif
+
+TEST(flatten_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_flatten_op<::anakin::saber::X86>(ctx, false);
+}
+
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
 USE_OP(reshape);
 USE_OP_ITSELF(flatten);
+USE_CPU_ANAKIN_CONVERTER(flatten);
+#ifdef PADDLE_WITH_CUDA
 USE_ANAKIN_CONVERTER(flatten);
+#endif

paddle/fluid/inference/anakin/convert/test_pool2d_op.cc

@@ -19,15 +19,14 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-void test_pool2d(bool global_pooling, bool ceil_mode,
+template <typename TargetT>
+void test_pool2d(const platform::DeviceContext& context, bool use_gpu,
+                 bool global_pooling, bool ceil_mode,
                  std::string pool_type = "max") {
-  auto* pool2d_converter =
-      Registry<AnakinOpConverter>::Global().Lookup("pool2d");
-  ASSERT_TRUE(pool2d_converter);
-
   framework::Scope scope;
   std::unordered_set<std::string> parameters;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
 
   // The ITensor's Dims should not contain the batch size.
   // So, the ITensor's Dims of input and output should be C * H * W.
@@ -64,56 +63,61 @@ void test_pool2d(bool global_pooling, bool ceil_mode,
   validator.Execute(1);
 }
 
-void test_pool2d2(bool global_pooling, bool ceil_mode,
-                  std::string pool_type = "max") {
-  auto* pool2d_converter =
-      Registry<AnakinOpConverter>::Global().Lookup("pool2d");
-  ASSERT_TRUE(pool2d_converter);
-
-  framework::Scope scope;
-  std::unordered_set<std::string> parameters;
-  AnakinConvertValidation validator(parameters, &scope);
-
-  // The ITensor's Dims should not contain the batch size.
-  // So, the ITensor's Dims of input and output should be C * H * W.
-  validator.DeclInputVar("pool2d_x", {1, 1, 17, 17});
-  validator.DeclOutputVar("pool2d_out", {1, 1, 17, 17});
-
-  // Prepare Op description
-  framework::OpDesc desc;
-  desc.SetType("pool2d");
-  desc.SetInput("X", {"pool2d_x"});
-  desc.SetOutput("Out", {"pool2d_out"});
-
-  std::vector<int> ksize({3, 3});
-  std::vector<int> strides({1, 1});
-  std::vector<int> paddings({1, 1});
-  std::string pooling_t = pool_type;
+#ifdef PADDLE_WITH_CUDA
+TEST(Pool2dOpConverter, normal) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_pool2d<::anakin::saber::NV>(ctx, true, false, false);
+}
+TEST(Pool2dOpConverter, test_global_pooling) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_pool2d<::anakin::saber::NV>(ctx, true, true, false);
+}
 
-  desc.SetAttr("pooling_type", pooling_t);
-  desc.SetAttr("ksize", ksize);
-  desc.SetAttr("strides", strides);
-  desc.SetAttr("paddings", paddings);
-  desc.SetAttr("global_pooling", global_pooling);
-  desc.SetAttr("ceil_mode", true);
+TEST(Pool2dOpConverter, max_ceil_test) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_pool2d<::anakin::saber::NV>(ctx, true, false, true);
+}
 
-  LOG(INFO) << "set OP";
-  validator.SetOp(*desc.Proto());
-  LOG(INFO) << "execute";
+TEST(Pool2dOpConverter, avg_ceil_test) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_pool2d<::anakin::saber::NV>(ctx, true, false, true, "avg");
+}
+#endif
 
-  validator.Execute(1);
+TEST(Pool2dOpConverter, normal_cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_pool2d<::anakin::saber::X86>(ctx, false, false, false);
+}
+TEST(Pool2dOpConverter, test_global_pooling_cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_pool2d<::anakin::saber::X86>(ctx, false, true, false);
 }
 
-TEST(Pool2dOpConverter, normal) { test_pool2d(false, false); }
-TEST(Pool2dOpConverter, test_global_pooling) { test_pool2d(true, false); }
+TEST(Pool2dOpConverter, max_ceil_test_cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_pool2d<::anakin::saber::X86>(ctx, false, false, true);
+}
 
-TEST(Pool2dOpConverter, max_ceil_test) { test_pool2d(false, true); }
-TEST(Pool2dOpConverter, avg_ceil_test) { test_pool2d(false, true, "avg"); }
-TEST(Pool2dOpConverter, avg_ceil_test2) { test_pool2d2(false, true, "avg"); }
+TEST(Pool2dOpConverter, avg_ceil_test_cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_pool2d<::anakin::saber::X86>(ctx, false, false, true, "avg");
+}
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
 USE_OP(pool2d);
+USE_CPU_ANAKIN_CONVERTER(pool2d);
+
+#ifdef PADDLE_WITH_CUDA
 USE_ANAKIN_CONVERTER(pool2d);
+#endif

paddle/fluid/inference/anakin/convert/test_relu_op.cc

@@ -21,12 +21,14 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-static void test_relu_op(const std::string &op_type) {
-  auto *converter = Registry<AnakinOpConverter>::Global().Lookup(op_type);
-  PADDLE_ENFORCE(converter != nullptr);
+template <typename TargetT>
+static void test_activation_op(const std::string& op_type,
+                               const platform::DeviceContext& context,
+                               bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
   validator.DeclInputVar("act-X", {10, 6, 1, 1});
   validator.DeclOutputVar("act-Out", {10, 6, 1, 1});
   framework::OpDesc desc;
@@ -44,14 +46,44 @@ static void test_relu_op(const std::string &op_type) {
   validator.Execute(5);
 }
 
-TEST(activation, relu) { test_relu_op("relu"); }
-TEST(activation, leaky_relu) { test_relu_op("leaky_relu"); }
+#ifdef PADDLE_WITH_CUDA
+TEST(relu_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_activation_op<::anakin::saber::NV>("relu", ctx, true);
+}
+
+TEST(leaky_relu_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_activation_op<::anakin::saber::NV>("leaky_relu", ctx, true);
+}
+#endif
+
+/* seems bug here
+TEST(relu_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_activation_op<::anakin::saber::X86>("relu", ctx, false);
+}
+
+TEST(leaky_relu_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_activation_op<::anakin::saber::X86>("leaky_relu", ctx, false);
+}
+*/
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
 USE_OP(relu);
-USE_ANAKIN_CONVERTER(relu);
 USE_OP(leaky_relu);
+USE_CPU_ANAKIN_CONVERTER(relu);
+USE_CPU_ANAKIN_CONVERTER(leaky_relu);
+
+#ifdef PADDLE_WITH_CUDA
+USE_ANAKIN_CONVERTER(relu);
 USE_ANAKIN_CONVERTER(leaky_relu);
+#endif

paddle/fluid/inference/anakin/convert/test_reshape_op.cc

@@ -20,12 +20,12 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-TEST(reshape, test) {
-  auto* converter = Registry<AnakinOpConverter>::Global().Lookup("reshape");
-  ASSERT_TRUE(converter);
+template <typename TargetT>
+void test_reshape1_op(const platform::DeviceContext& context, bool use_gpu) {
   framework::Scope scope;
   std::unordered_set<std::string> parameters;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
 
   // validator.DeclInputVar("reshape-X", {2, 3, 3, 1});
   // validator.DeclOutputVar("reshape-Out", {3, 2, 1, 3});
@@ -45,10 +45,12 @@ TEST(reshape, test) {
   validator.Execute(1);
 }
 
-TEST(reshape, test2) {
+template <typename TargetT>
+void test_reshape2_op(const platform::DeviceContext& context, bool use_gpu) {
   framework::Scope scope;
   std::unordered_set<std::string> parameters;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
 
   validator.DeclInputVar("reshape-X", {1, 2, 4});
   validator.DeclOutputVar("reshape-Out", {1, 4, 2});
@@ -66,9 +68,39 @@ TEST(reshape, test2) {
   validator.Execute(1);
 }
 
+#ifdef PADDLE_WITH_CUDA
+TEST(reshape1_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_reshape1_op<::anakin::saber::NV>(ctx, true);
+}
+
+TEST(reshape2_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_reshape2_op<::anakin::saber::NV>(ctx, true);
+}
+#endif
+
+TEST(reshape1_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_reshape2_op<::anakin::saber::X86>(ctx, false);
+}
+
+TEST(reshape2_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_reshape2_op<::anakin::saber::X86>(ctx, false);
+}
+
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
 USE_OP(reshape);
+USE_CPU_ANAKIN_CONVERTER(reshape);
+
+#ifdef PADDLE_WITH_CUDA
 USE_ANAKIN_CONVERTER(reshape);
+#endif

paddle/fluid/inference/anakin/convert/test_softmax_op.cc

@@ -20,12 +20,12 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-TEST(softmax, test) {
-  auto* converter = Registry<AnakinOpConverter>::Global().Lookup("softmax");
-  ASSERT_TRUE(converter);
+template <typename TargetT>
+void test_softmax_op(const platform::DeviceContext& context, bool use_gpu) {
   framework::Scope scope;
   std::unordered_set<std::string> parameters;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
 
   validator.DeclInputVar("softmax-X", {1, 10, 2});
   validator.DeclOutputVar("softmax-Out", {1, 10, 2});
@@ -41,9 +41,27 @@ TEST(softmax, test) {
   validator.Execute(1);
 }
 
+#ifdef PADDLE_WITH_CUDA
+TEST(softmax_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_softmax_op<::anakin::saber::NV>(ctx, true);
+}
+#endif
+
+TEST(relu_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_softmax_op<::anakin::saber::X86>(ctx, false);
+}
+
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
 USE_OP(softmax);
+USE_CPU_ANAKIN_CONVERTER(softmax);
+
+#ifdef PADDLE_WITH_CUDA
 USE_ANAKIN_CONVERTER(softmax);
+#endif

paddle/fluid/inference/anakin/convert/test_split_op.cc

@@ -21,12 +21,14 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-template <int Axis>
-void AnakinSliceTest(const std::vector<int> &in_shape,
+template <typename TargetT, int Axis>
+void AnakinSliceTest(const platform::DeviceContext &context, bool use_gpu,
+                     const std::vector<int> &in_shape,
                      const std::vector<int> &sections) {
   std::unordered_set<std::string> parameters({""});
   framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
 
   validator.DeclInputVar("split_input", in_shape);
   std::vector<std::string> output_vars;
@@ -55,51 +57,58 @@ void AnakinSliceTest(const std::vector<int> &in_shape,
 
 // batch = 0, axis = 1, same shape
 TEST(split_op, test_same_shape_axis1_batch1) {
-  AnakinSliceTest<1>({1, 4, 2, 2}, {2, 2});
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  AnakinSliceTest<::anakin::saber::NV, 1>(ctx, true, {1, 4, 2, 2}, {2, 2});
 }
 // batch = 0, axis = 1, different shape
 TEST(split_op, test_different_shape_axis1_batch1) {
-  AnakinSliceTest<1>({1, 3, 2, 2}, {2, 1});
-}
-// batch = 10, axis = 1, same shape
-TEST(split_op, test_same_shape_axis1_batch10) {
-  AnakinSliceTest<1>({1, 4, 2, 2}, {2, 2});
-}
-// batch = 10, axis = 1, different shape
-TEST(split_op, test_different_shape_axis1_batch10) {
-  AnakinSliceTest<1>({1, 3, 2, 2}, {2, 1});
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  AnakinSliceTest<::anakin::saber::NV, 1>(ctx, true, {1, 3, 2, 2}, {2, 1});
 }
 // batch = 0, axis = 2, same shape
 TEST(split_op, test_same_shape_axis2_batch1) {
-  AnakinSliceTest<2>({1, 3, 4, 2}, {2, 2});
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  AnakinSliceTest<::anakin::saber::NV, 2>(ctx, true, {1, 3, 4, 2}, {2, 2});
 }
 // batch = 0, axis = 2, different shape
 TEST(split_op, test_different_shape_axis2_batch1) {
-  AnakinSliceTest<2>({1, 3, 3, 2}, {2, 1});
-}
-// batch = 10, axis = 2, same shape
-TEST(split_op, test_same_shape_axis2_batch10) {
-  AnakinSliceTest<2>({1, 3, 4, 2}, {2, 2});
-}
-// batch = 10, axis = 2, different shape
-TEST(split_op, test_different_shape_axis2_batch10) {
-  AnakinSliceTest<2>({1, 3, 3, 2}, {2, 1});
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  AnakinSliceTest<::anakin::saber::NV, 2>(ctx, true, {1, 3, 3, 2}, {2, 1});
 }
+
 // batch = 0, axis = 3, same shape
 TEST(split_op, test_same_shape_axis3_batch1) {
-  AnakinSliceTest<3>({1, 3, 2, 4}, {2, 2});
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  AnakinSliceTest<::anakin::saber::NV, 3>(ctx, true, {1, 3, 2, 4}, {2, 2});
 }
 // batch = 0, axis = 3, different shape
 TEST(split_op, test_different_shape_axis3_batch1) {
-  AnakinSliceTest<3>({1, 3, 2, 3}, {2, 1});
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  AnakinSliceTest<::anakin::saber::NV, 3>(ctx, true, {1, 3, 2, 3}, {2, 1});
 }
-// batch = 10, axis = 3, same shape
-TEST(split_op, test_same_shape_axis3_batch10) {
-  AnakinSliceTest<3>({1, 3, 2, 4}, {2, 2});
+
+TEST(split_op, test_different_shape_axis1_batch1_cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  AnakinSliceTest<::anakin::saber::X86, 1>(ctx, false, {1, 3, 2, 3}, {2, 1});
+}
+
+TEST(split_op, test_different_shape_axis2_batch1_cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  AnakinSliceTest<::anakin::saber::X86, 2>(ctx, false, {1, 3, 4, 2}, {2, 2});
 }
-// batch = 10, axis = 3, different shape
-TEST(split_op, test_different_shape_axis3_batch10) {
-  AnakinSliceTest<3>({1, 3, 2, 3}, {2, 1});
+
+TEST(split_op, test_different_shape_axis3_batch1_cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  AnakinSliceTest<::anakin::saber::X86, 3>(ctx, false, {1, 3, 2, 4}, {2, 2});
 }
 
 }  // namespace anakin
@@ -107,4 +116,7 @@ TEST(split_op, test_different_shape_axis3_batch10) {
 }  // namespace paddle
 
 USE_OP(split);
+USE_CPU_ANAKIN_CONVERTER(split);
+#ifdef PADDLE_WITH_CUDA
 USE_ANAKIN_CONVERTER(split);
+#endif

paddle/fluid/inference/anakin/convert/test_sum_op.cc

@@ -22,10 +22,12 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-TEST(sum, native) {
+template <typename TargetT>
+static void test_sum_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
   validator.DeclInputVar("sum_x1", {1, 2, 1, 2});
   validator.DeclInputVar("sum_x2", {1, 2, 1, 2});
   validator.DeclOutputVar("sum_out", {1, 2, 1, 2});
@@ -40,9 +42,26 @@ TEST(sum, native) {
   validator.Execute(1);
 }
 
+#ifdef PADDLE_WITH_CUDA
+TEST(sum_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_sum_op<::anakin::saber::NV>(ctx, true);
+}
+#endif
+
+TEST(sum_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_sum_op<::anakin::saber::X86>(ctx, false);
+}
+
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
 USE_OP(sum);
+USE_CPU_ANAKIN_CONVERTER(sum);
+#ifdef PADDLE_WITH_CUDA
 USE_ANAKIN_CONVERTER(sum);
+#endif

paddle/fluid/inference/anakin/convert/test_transpose_op.cc

@@ -20,12 +20,12 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-TEST(transpose_op, test) {
-  auto* converter = Registry<AnakinOpConverter>::Global().Lookup("transpose");
-  ASSERT_TRUE(converter != nullptr);
+template <typename TargetT>
+void test_transpose1_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
   validator.DeclInputVar("transpose-X", {2, 3, 4, 5});
   validator.DeclOutputVar("transpose-Out", {4, 2, 5, 3});
 
@@ -43,11 +43,12 @@ TEST(transpose_op, test) {
   validator.Execute(3);
 }
 
-// test input shape's dims < 4
-TEST(transpose_op, test2) {
+template <typename TargetT>
+void test_transpose2_op(const platform::DeviceContext& context, bool use_gpu) {
   std::unordered_set<std::string> parameters;
   framework::Scope scope;
-  AnakinConvertValidation validator(parameters, &scope);
+  AnakinConvertValidation<TargetT> validator(parameters, &scope, context,
+                                             use_gpu);
   validator.DeclInputVar("transpose-X", {3, 4, 5});
   validator.DeclOutputVar("transpose-Out", {3, 5, 4});
 
@@ -65,9 +66,38 @@ TEST(transpose_op, test2) {
   validator.Execute(1);
 }
 
+#ifdef PADDLE_WITH_CUDA
+TEST(transpose1_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_transpose1_op<::anakin::saber::NV>(ctx, true);
+}
+
+TEST(transpose2_op, gpu) {
+  platform::CUDAPlace gpu_place(0);
+  platform::CUDADeviceContext ctx(gpu_place);
+  test_transpose2_op<::anakin::saber::NV>(ctx, true);
+}
+#endif
+
+TEST(transpose1_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_transpose2_op<::anakin::saber::X86>(ctx, false);
+}
+
+TEST(transpose2_op, cpu) {
+  platform::CPUPlace cpu_place;
+  platform::CPUDeviceContext ctx(cpu_place);
+  test_transpose2_op<::anakin::saber::X86>(ctx, false);
+}
+
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
 USE_OP(transpose);
+USE_CPU_ANAKIN_CONVERTER(transpose);
+#ifdef PADDLE_WITH_CUDA
 USE_ANAKIN_CONVERTER(transpose);
+#endif

paddle/fluid/inference/anakin/convert/transpose.cc

@@ -17,20 +17,16 @@
 #include <string>
 #include <vector>
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::saber::NV;
-using anakin::saber::Shape;
 using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
 namespace anakin {
 
-void TransposeOpConverter::operator()(const framework::proto::OpDesc &op,
-                                      const framework::BlockDesc &block_desc,
-                                      const framework::Scope &scope,
-                                      bool test_mode) {
+template <typename TargetT>
+void TransposeOpConverter<TargetT>::operator()(
+    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
+    const framework::Scope &scope, bool test_mode) {
   framework::OpDesc op_desc(op, nullptr);
   PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
   PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
@@ -38,7 +34,7 @@ void TransposeOpConverter::operator()(const framework::proto::OpDesc &op,
   auto input = op_desc.Input("X").front();
   auto output = op_desc.Output("Out").front();
   auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
-  engine_->AddOp(op_name, "Permute", {input}, {output});
+  this->engine_->AddOp(op_name, "Permute", {input}, {output});
 
   auto axis = boost::get<std::vector<int>>(op_desc.GetAttr("axis"));
   size_t axis_size = axis.size();
@@ -46,11 +42,17 @@ void TransposeOpConverter::operator()(const framework::proto::OpDesc &op,
     axis.push_back(axis_size);
     axis_size += 1;
   }
-  engine_->AddOpAttr<PTuple<int>>(op_name, "dims", axis);
+  this->engine_->template AddOpAttr<PTuple<int>>(op_name, "dims", axis);
 }
 
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_ANAKIN_OP_CONVERTER(transpose, TransposeOpConverter);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_CUDA_ANAKIN_OP_CONVERTER(transpose,
+                                  TransposeOpConverter<::anakin::saber::NV>);
+#endif
+
+REGISTER_CPU_ANAKIN_OP_CONVERTER(transpose,
+                                 TransposeOpConverter<::anakin::saber::X86>);

paddle/fluid/inference/anakin/convert/transpose.h

@@ -20,7 +20,8 @@ namespace paddle {
 namespace inference {
 namespace anakin {
 
-class TransposeOpConverter : public AnakinOpConverter {
+template <typename TargetT>
+class TransposeOpConverter : public AnakinOpConverter<TargetT> {
  public:
   TransposeOpConverter() = default;
 

paddle/fluid/inference/anakin/convert/ut_helper.h

@@ -32,14 +32,8 @@ limitations under the License. */
 #include "paddle/fluid/inference/utils/singleton.h"
 #include "paddle/fluid/platform/enforce.h"
 
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
 using anakin::Precision;
-using anakin::saber::NV;
 using anakin::saber::X86;
-using anakin::saber::Shape;
-using anakin::PBlock;
-using anakin::PTuple;
 
 namespace paddle {
 namespace inference {
@@ -55,8 +49,8 @@ float random(float low, float high) {
   return dist(mt);
 }
 
-void RandomizeTensor(framework::LoDTensor* tensor, const platform::Place& place,
-                     const platform::DeviceContext& ctx) {
+void RandomizeTensor(framework::LoDTensor* tensor,
+                     const platform::Place& place) {
   auto dims = tensor->dims();
   size_t num_elements = analysis::AccuDims(dims, dims.size());
   PADDLE_ENFORCE_GT(num_elements, 0);
@@ -78,17 +72,19 @@ void RandomizeTensor(framework::LoDTensor* tensor, const platform::Place& place,
  * anakin
  * layer.
  */
+template <typename TargetT>
 class AnakinConvertValidation {
-  using AnakinNvEngineT = AnakinEngine<NV, Precision::FP32>;
+  using AnakinNvEngineT = AnakinEngine<TargetT, Precision::FP32>;
 
  public:
   AnakinConvertValidation() = delete;
 
   AnakinConvertValidation(const std::unordered_set<std::string>& parameters,
-                          framework::Scope* scope)
-      : parameters_(parameters), scope_(scope), place_(0) {
-    PADDLE_ENFORCE_EQ(cudaStreamCreate(&stream_), 0);
-    engine_.reset(new AnakinEngine<NV, Precision::FP32>(true));
+                          framework::Scope* scope,
+                          const platform::DeviceContext& ctx,
+                          bool use_gpu = true)
+      : parameters_(parameters), scope_(scope), ctx_(ctx), use_gpu_(use_gpu) {
+    engine_.reset(new AnakinEngine<TargetT, Precision::FP32>(true));
   }
 
   // Declare a Variable as input with random initialization.
@@ -108,11 +104,10 @@ class AnakinConvertValidation {
   }
 
   void DeclVar(const std::string& name, const std::vector<int> dim_vec) {
-    platform::CUDADeviceContext ctx(place_);
     auto* x = scope_->Var(name);
     auto* x_tensor = x->GetMutable<framework::LoDTensor>();
     x_tensor->Resize(framework::make_ddim(dim_vec));
-    RandomizeTensor(x_tensor, place_, ctx);
+    RandomizeTensor(x_tensor, ctx_.GetPlace());
 
     std::vector<int64_t> dim_vec_int64;
     for (auto& ele : dim_vec) {
@@ -132,7 +127,7 @@ class AnakinConvertValidation {
     // should init anakin engine here.
 
     auto& block_desc = program_desc_.Block(framework::kRootBlockIndex);
-    Singleton<AnakinOpConverter>::Global().ConvertOp(
+    Singleton<AnakinOpConverter<TargetT>>::Global().ConvertOp(
         desc, block_desc, parameters_, *scope_, engine_.get(),
         true /*test_mode*/);
     engine_->Freeze();
@@ -160,11 +155,8 @@ class AnakinConvertValidation {
   void Execute(int batch_size,
                std::unordered_set<std::string> neglected_output = {}) {
     // Execute Fluid Op
-    platform::CUDADeviceContext ctx(place_);
-    op_->Run(*scope_, place_);
+    op_->Run(*scope_, ctx_.GetPlace());
 
-    // std::vector<framework::LoDTensor> input_vector;
-    // std::vector<framework::LoDTensor> output_vector;
     std::map<std::string, framework::LoDTensor*> inputs;
     for (const auto& input : op_desc_->InputArgumentNames()) {
       if (parameters_.count(input)) continue;
@@ -180,20 +172,27 @@ class AnakinConvertValidation {
       std::vector<float> fluid_out;
       auto* var = scope_->FindVar(output);
       auto tensor = var->GetMutable<framework::LoDTensor>();
-      framework::TensorToVector(*tensor, ctx, &fluid_out);
+      framework::TensorToVector(*tensor, ctx_, &fluid_out);
       fluid_outputs.push_back(fluid_out);
 
       outputs.insert({output, tensor});
     }
 
-    engine_->Execute(inputs, outputs, stream_);
+    if (!use_gpu_) {
+      engine_->Execute(inputs, outputs);
+    } else {
+      cudaStream_t stream;
+      PADDLE_ENFORCE_EQ(cudaStreamCreate(&stream), 0);
+      engine_->Execute(inputs, outputs, stream);
+    }
+
     int i_output = 0;
     for (const auto& output : op_desc_->OutputArgumentNames()) {
       if (neglected_output.count(output)) continue;
       std::vector<float> anakin_out;
       auto* var = scope_->FindVar(output);
       auto tensor = var->GetMutable<framework::LoDTensor>();
-      framework::TensorToVector(*tensor, ctx, &anakin_out);
+      framework::TensorToVector(*tensor, ctx_, &anakin_out);
 
       size_t anakin_out_size = anakin_out.size();
       auto fluid_out = fluid_outputs[i_output++];
@@ -205,15 +204,17 @@ class AnakinConvertValidation {
 
  private:
   std::unique_ptr<AnakinNvEngineT> engine_{nullptr};
-  cudaStream_t stream_;
   std::unique_ptr<framework::OperatorBase> op_;
   std::unique_ptr<framework::OpDesc> op_desc_;
   framework::ProgramDesc program_desc_;
   const std::unordered_set<std::string>& parameters_;
   framework::Scope* scope_;
-  platform::CUDAPlace place_;
+  const platform::DeviceContext& ctx_;
+  bool use_gpu_{true};
 };
 
+template class AnakinConvertValidation<::anakin::saber::NV>;
+template class AnakinConvertValidation<::anakin::saber::X86>;
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/anakin/engine.cc

@@ -69,11 +69,11 @@ void AnakinEngine<TargetT, PrecisionType, RunType>::AddOp(
 }
 
 template <typename TargetT, Precision PrecisionType, OpRunType RunType>
-void AnakinEngine<TargetT, PrecisionType, RunType>::Execute(
-    const std::map<std::string, framework::LoDTensor *> &inputs,
-    const std::map<std::string, framework::LoDTensor *> &outputs,
-    cudaStream_t stream) {
+void AnakinEngine<TargetT, PrecisionType, RunType>::BindInput(
+    const std::map<std::string, framework::LoDTensor *> &inputs) {
+#ifdef PADDLE_WITH_CUDA
   cudaDeviceSynchronize();
+#endif
   for (const auto &input : inputs) {
     auto *tensor = input.second;
     auto *data = tensor->data<float>();
@@ -105,6 +105,35 @@ void AnakinEngine<TargetT, PrecisionType, RunType>::Execute(
                                                        fluid_input_shape);
     anakin_input->copy_from(tmp_anakin_tensor);
   }
+}
+
+template <typename TargetT, Precision PrecisionType, OpRunType RunType>
+void AnakinEngine<TargetT, PrecisionType, RunType>::Execute(
+    const std::map<std::string, framework::LoDTensor *> &inputs,
+    const std::map<std::string, framework::LoDTensor *> &outputs) {
+  BindInput(inputs);
+  net_->prediction();
+  for (const auto &output : outputs) {
+    platform::CPUPlace cpu_place;
+    auto *tensor = output.second;
+    auto *anakin_output = net_->get_out(output.first);
+    auto *anakin_data = anakin_output->data();
+    auto anakin_output_shape = anakin_output->valid_shape();
+    tensor->Resize(framework::make_ddim(anakin_output_shape));
+    auto *fluid_data = tensor->mutable_data<float>(cpu_place);
+    memory::Copy(cpu_place, static_cast<void *>(fluid_data), cpu_place,
+                 static_cast<void *>(anakin_data),
+                 tensor->numel() * sizeof(float));
+  }
+}
+
+#ifdef PADDLE_WITH_CUDA
+template <typename TargetT, Precision PrecisionType, OpRunType RunType>
+void AnakinEngine<TargetT, PrecisionType, RunType>::Execute(
+    const std::map<std::string, framework::LoDTensor *> &inputs,
+    const std::map<std::string, framework::LoDTensor *> &outputs,
+    cudaStream_t stream) {
+  BindInput(inputs);
   net_->prediction();
   cudaDeviceSynchronize();
   for (const auto &output : outputs) {
@@ -121,6 +150,7 @@ void AnakinEngine<TargetT, PrecisionType, RunType>::Execute(
   }
   cudaDeviceSynchronize();
 }
+#endif
 
 template <typename TargetT, Precision PrecisionType, OpRunType RunType>
 void AnakinEngine<TargetT, PrecisionType, RunType>::Freeze() {
@@ -140,7 +170,15 @@ AnakinEngine<TargetT, PrecisionType, RunType>::Clone() {
   return std::unique_ptr<AnakinEngine>(engine);
 }
 
+#ifdef PADDLE_WITH_CUDA
 template class AnakinEngine<::anakin::saber::NV, ::anakin::Precision::FP32>;
+template class AnakinEngineManager<::anakin::saber::NV>;
+#endif
+
+template class AnakinEngine<::anakin::saber::X86, ::anakin::Precision::FP32>;
+template class AnakinEngineManager<::anakin::saber::X86>;
+
+// template class AnakinEngine<::anakin::saber::X86, ::anakin::Precision::FP32>;
 }  // namespace anakin
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/anakin/engine.h

@@ -32,7 +32,6 @@
 #include "saber/saber_types.h"
 
 using anakin::Precision;
-using anakin::saber::NV;
 
 namespace anakin {
 
@@ -94,9 +93,16 @@ class AnakinEngine {
   void Save(std::string path) { graph_->save(path); }
   bool IsInit() { return initialized_; }
   int GetDevice() { return device_; }
+  void Execute(const std::map<std::string, framework::LoDTensor *> &inputs,
+               const std::map<std::string, framework::LoDTensor *> &outputs);
+#ifdef PADDLE_WITH_CUDA
   void Execute(const std::map<std::string, framework::LoDTensor *> &inputs,
                const std::map<std::string, framework::LoDTensor *> &outputs,
                cudaStream_t stream);
+#endif
+
+ private:
+  void BindInput(const std::map<std::string, framework::LoDTensor *> &inputs);
 
  private:
   bool initialized_{false};
@@ -108,24 +114,25 @@ class AnakinEngine {
   std::vector<std::string> program_inputs_;
 };
 
+template <typename TargetT>
 class AnakinEngineManager {
-  using AnakinNvEngineT = AnakinEngine<NV, Precision::FP32>;
+  using AnakinEngineT = AnakinEngine<TargetT, Precision::FP32>;
 
  public:
   bool HasEngine(const std::string &name) const {
     if (engines_.count(name) == 0) return false;
     return engines_.at(name).get() != nullptr;
   }
-  AnakinNvEngineT *Get(const std::string &name) const {
+  AnakinEngineT *Get(const std::string &name) const {
     return engines_.at(name).get();
   }
 
-  AnakinNvEngineT *Create(
-      bool need_summary, int device, int max_batch_size,
-      std::map<std::string, std::vector<int>> max_input_shape,
-      std::vector<std::string> program_inputs, std::string engine_name) {
+  AnakinEngineT *Create(bool need_summary, int device, int max_batch_size,
+                        std::map<std::string, std::vector<int>> max_input_shape,
+                        std::vector<std::string> program_inputs,
+                        std::string engine_name) {
     std::unique_lock<std::mutex> lk(mut_);
-    auto *p = new AnakinEngine<NV, Precision::FP32>(
+    auto *p = new AnakinEngine<TargetT, Precision::FP32>(
         need_summary, device, max_batch_size, max_input_shape, program_inputs);
     engines_[engine_name].reset(p);
     return p;
@@ -138,7 +145,7 @@ class AnakinEngineManager {
   }
 
  private:
-  std::unordered_map<std::string, std::unique_ptr<AnakinNvEngineT>> engines_;
+  std::unordered_map<std::string, std::unique_ptr<AnakinEngineT>> engines_;
   std::mutex mut_;
 };
 }  // namespace anakin

paddle/fluid/inference/analysis/argument.h

@@ -64,20 +64,20 @@ struct Argument {
 
   bool Has(const std::string& key) const { return valid_fields_.count(key); }
 
-#define DECL_ARGUMENT_FIELD(field__, Field, type__) \
- public:                                            \
-  type__& field__() {                               \
-    PADDLE_ENFORCE(Has(#field__));                  \
-    return field__##_;                              \
-  }                                                 \
-  void Set##Field(const type__& x) {                \
-    field__##_ = x;                                 \
-    valid_fields_.insert(#field__);                 \
-  }                                                 \
-  DECL_ARGUMENT_FIELD_VALID(field__);               \
-  type__* field__##_ptr() { return &field__##_; }   \
-                                                    \
- private:                                           \
+#define DECL_ARGUMENT_FIELD(field__, Field, type__)          \
+ public:                                                     \
+  type__& field__() {                                        \
+    PADDLE_ENFORCE(Has(#field__), "There is no such field"); \
+    return field__##_;                                       \
+  }                                                          \
+  void Set##Field(const type__& x) {                         \
+    field__##_ = x;                                          \
+    valid_fields_.insert(#field__);                          \
+  }                                                          \
+  DECL_ARGUMENT_FIELD_VALID(field__);                        \
+  type__* field__##_ptr() { return &field__##_; }            \
+                                                             \
+ private:                                                    \
   type__ field__##_;
 
 #define DECL_ARGUMENT_FIELD_VALID(field__) \

paddle/fluid/inference/analysis/ir_pass_manager.cc

@@ -114,6 +114,7 @@ void IRPassManager::CreatePasses(Argument *argument,
     if (pass_name == "anakin_subgraph_pass") {
       pass->Set("program",
                 new framework::ProgramDesc *(&argument->main_program()));
+      pass->Set("use_gpu", new bool(argument->use_gpu()));
       pass->Set("gpu_device_id", new int(argument->gpu_device_id()));
       pass->Set("model_from_memory", new bool(argument->model_from_memory()));
       pass->Set("engine_opt_info", new std::map<std::string, std::string>(

paddle/fluid/inference/analysis/ir_passes/anakin_subgraph_pass.cc

@@ -194,20 +194,49 @@ void AnakinSubgraphPass::CreateAnakinOp(
   auto max_batch_size = Get<int>("max_batch_size");
   auto program_inputs = program_desc->GetFeedTargetNames();
 
-  auto *anakin_engine =
-      inference::Singleton<anakin::AnakinEngineManager>::Global().Create(
-          true, Get<int>("gpu_device_id"), max_batch_size, max_input_shape,
-          program_inputs, engine_key);
+  bool use_gpu = Get<bool>("use_gpu");
+  SetAttr(op_desc->Proto(), "use_gpu", use_gpu);
+
+  if (use_gpu) {
+#ifdef PADDLE_WITH_CUDA
+    inference::Singleton<
+        anakin::AnakinEngineManager<::anakin::saber::NV>>::Global()
+        .Create(true, Get<int>("gpu_device_id"), max_batch_size,
+                max_input_shape, program_inputs, engine_key);
+#endif
+  } else {
+    inference::Singleton<
+        anakin::AnakinEngineManager<::anakin::saber::X86>>::Global()
+        .Create(true, Get<int>("gpu_device_id"), max_batch_size,
+                max_input_shape, program_inputs, engine_key);
+  }
 
   auto *scope = param_scope();
   std::unordered_set<std::string> param_set(params.begin(), params.end());
   framework::BlockDesc block_desc_temp(nullptr, block_desc.Proto());
-
-  inference::Singleton<inference::anakin::AnakinOpConverter>::Global()
-      .ConvertBlockToAnakinEngine(
-          &block_desc_temp, scope,
-          std::vector<std::string>(input_names.begin(), input_names.end()),
-          param_set, output_mapping, anakin_engine);
+  if (use_gpu) {
+    auto *anakin_engine =
+        inference::Singleton<inference::anakin::AnakinEngineManager<
+            ::anakin::saber::NV>>::Global()
+            .Get(engine_key);
+    inference::Singleton<
+        inference::anakin::AnakinOpConverter<::anakin::saber::NV>>::Global()
+        .ConvertBlockToAnakinEngine(
+            &block_desc_temp, scope,
+            std::vector<std::string>(input_names.begin(), input_names.end()),
+            param_set, output_mapping, anakin_engine);
+  } else {
+    auto *anakin_engine =
+        inference::Singleton<inference::anakin::AnakinEngineManager<
+            ::anakin::saber::X86>>::Global()
+            .Get(engine_key);
+    inference::Singleton<
+        inference::anakin::AnakinOpConverter<::anakin::saber::X86>>::Global()
+        .ConvertBlockToAnakinEngine(
+            &block_desc_temp, scope,
+            std::vector<std::string>(input_names.begin(), input_names.end()),
+            param_set, output_mapping, anakin_engine);
+  }
 }
 
 }  // namespace analysis

paddle/fluid/inference/api/CMakeLists.txt

@@ -70,4 +70,3 @@ if (WITH_ANAKIN AND WITH_MKL) # only needed in CI
     anakin_target(inference_anakin_api)
     anakin_target(inference_anakin_api_shared)
 endif()
-inference_analysis_test(faster_rcnn_test SRCS faster_rcnn_test.cc EXTRA_DEPS paddle_fluid)

paddle/fluid/inference/api/analysis_config.cc

@@ -268,9 +268,11 @@ void AnalysisConfig::Update() {
     PADDLE_ENFORCE(!use_tensorrt_,
                    "Anakin sub-graph and TensorRT sub-graph are not allowed to "
                    "run at the same time!");
-    PADDLE_ENFORCE(
-        use_gpu_,
-        "Anakin sub-graph engine need gpu, please use the EnableGpu API.");
+    if (use_gpu_) {
+      LOG(INFO) << "Run Anakin GPU mode";
+    } else {
+      LOG(INFO) << "Run Anakin CPU mode";
+    }
 
     pass_builder()->ClearPasses();
     for (const auto &pass : kAnakinSubgraphPasses) {

paddle/fluid/inference/api/analysis_predictor.cc

@@ -382,7 +382,7 @@ void AnalysisPredictor::PrepareArgument() {
     argument_.SetTensorRtUseStaticEngine(config_.trt_use_static_engine_);
   }
 
-  if (config_.use_gpu() && config_.anakin_engine_enabled()) {
+  if (config_.anakin_engine_enabled()) {
     argument_.SetAnakinMaxBatchSize(config_.anakin_max_batchsize_);
     argument_.SetAnakinMaxInputShape(config_.anakin_max_input_shape_);
     argument_.SetAnakinMinSubgraphSize(config_.anakin_min_subgraph_size_);

paddle/fluid/operators/anakin/anakin_engine_op.h

@@ -34,28 +34,16 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-using FluidDT = framework::proto::VarType_Type;
 using inference::Singleton;
-
-using anakin::graph::GraphGlobalMem;
-using anakin::AK_FLOAT;
-using anakin::Precision;
-using anakin::saber::NV;
-using anakin::saber::X86;
-using anakin::saber::Shape;
-using anakin::PBlock;
-using anakin::PTuple;
 using inference::anakin::AnakinEngine;
 
 class AnakinEngineOp : public framework::OperatorBase {
-  using AnakinNvEngineT = AnakinEngine<NV, Precision::FP32>;
-
  private:
   std::vector<std::string> input_names_;
   std::unordered_set<std::string> param_names_;
-  mutable AnakinNvEngineT *anakin_engine_;
   std::string engine_key_;
   std::string engine_serialized_data_;
+  bool use_gpu_;
 
  public:
   AnakinEngineOp(const std::string &type,
@@ -66,10 +54,10 @@ class AnakinEngineOp : public framework::OperatorBase {
     input_names_ = Inputs("Xs");
     engine_key_ = Attr<std::string>("engine_key");
     auto params = Attr<std::vector<std::string>>("parameters");
+    use_gpu_ = Attr<bool>("use_gpu");
     for (const auto &param : params) {
       param_names_.insert(param);
     }
-    anakin_engine_ = nullptr;
   }
 
  protected:
@@ -80,7 +68,6 @@ class AnakinEngineOp : public framework::OperatorBase {
 
   void RunAnakin(const framework::Scope &scope,
                  const platform::Place &dev_place) const {
-    auto *engine = GetEngine(scope, dev_place);
     platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
     auto &dev_ctx = *pool.Get(dev_place);
     auto stream =
@@ -92,7 +79,6 @@ class AnakinEngineOp : public framework::OperatorBase {
         Attr<std::vector<std::string>>("output_name_mapping");
 
     std::map<std::string, framework::LoDTensor *> inputs;
-    // Convert input tensor from fluid to engine.
     for (const auto &x : Inputs("Xs")) {
       if (param_names_.count(x)) continue;
       auto &t =
@@ -110,17 +96,21 @@ class AnakinEngineOp : public framework::OperatorBase {
       outputs.insert({output_maps[output_index], fluid_t});
       output_index += 1;
     }
-    engine->Execute(inputs, outputs, stream);
-  }
-
-  AnakinNvEngineT *GetEngine(const framework::Scope &scope,
-                             const platform::Place &dev_place) const {
-    if (anakin_engine_ == nullptr) {
-      anakin_engine_ =
-          inference::Singleton<inference::anakin::AnakinEngineManager>::Global()
+    if (use_gpu_) {
+#ifdef PADDLE_WITH_CUDA
+      auto *engine =
+          inference::Singleton<inference::anakin::AnakinEngineManager<
+              ::anakin::saber::NV>>::Global()
+              .Get(engine_key_);
+      engine->Execute(inputs, outputs, stream);
+#endif
+    } else {
+      auto *engine =
+          inference::Singleton<inference::anakin::AnakinEngineManager<
+              ::anakin::saber::X86>>::Global()
               .Get(engine_key_);
+      engine->Execute(inputs, outputs);
     }
-    return anakin_engine_;
   }
 };